This is rehearsal for Adarsh Vishwakarma. Round 4 only. R1 was 13 Aug 2026. R2 was 18 Aug 2026. R3 is done. Job 10454435 still has no public write-up that names a live-round question. Unnamed stays unnamed. Other people’s leftover-hour stems are shape evidence, not a list for this Job ID. One prompt, one spoken story, one project. Live Code is Java; that split lives on the home page — say it once if they ask, then stop.
How to use this page
Each leftover prompt has one first story. If that project already appeared in the hour, take the backup in the next table — a second story, not another trophy on the same intern. Spoken answers stay on one product. Follow-ups unpack that product. Full Leadership Principle STARs live in chapter 3.
What they might ask
A named leftover-hour stem other candidates wrote down.
A Leadership Principle by name, with no extra sentence.
Current role, architecture, or “tell me about a time…” after the intro.
Labels: IE-asked = another leftover hour named this stem (or listed it as a topic). Resume-derived = speakable from the Aug 2026 PDF even when the exact interviewer sentence was not found. GFG April 2026 bullets sit under a page heading “Common Questions” — that is not the same as “I was asked.” GFG 2025 numbered the OA as Round 1; their Bar Raiser after two lives is leftover-hour shape, not your Round 3.
What you say
I take the prompt they just asked. I tell one story from one product, stop on the resume number, and let them probe the same path. The full unique list with click-through is 9. LP list. Non-LP stems (why Amazon, live issues, questions for them) are 10. Behavioral.
They might ask next
If they ask
You answer
What did you implement?
The module on that story: RLS and Redis at Ylogx, hybrid retrieval and traces at IQVIA, GStreamer / ZED / costmap software at Horizon, YOLOv9 gate at Argus. Placement 17th is we.
They name a principle instead of a leftover stem
Use the spread on Home and the spoken STAR in chapter 3. This page is leftover wording mapped onto those same stories.
They paste code after an LP-only invite
Write Java. Other loops were told LP-only and still got DSA (LC 7406809, LC 6806195). Shape evidence, not a forecast.
Keep straight
One spoken answer, one project. Cross-product notes live in Keep straight, not inside What you say.
IQVIA is current (Apr 2026–present). Ylogx ran Nov 2024–Oct 2025. Horizon was Feb–Jun 2024 for ERC. Dates, not a contest.
99.9% uptime is Ylogx. IQVIA has no uptime or latency percent on the resume.
Tighten
Stop after the metric. If they want another prompt, start the next card. Do not list StratifyLabs, GiftedBooks, and Argus onto an intern walkthrough unless they pick a project.
If they already heard a project
Several leftover hours asked more than one named stem. Speak one project per answer. If they already heard project X, use the backup on that Leadership Principle — a second story, not the same intern again.
What they might ask
A second leftover stem after you already told Ylogx, IQVIA, Horizon, or Argus.
The same principle again with different wording (challenge, simplify, POCs).
Hours that named several stems in one sitting (other loops): LC 6806195 (feedback, critical issues, beyond task); LC 8362604 (challenge, deadline, identified, simplifying); LC 8430457 (disagreement, critical feedback, exceeded expectations); LC 6866029 (failed, missed a deadline, current architecture); LC 7035289 (not good at, learn a new technology, most challenging project); LC 7850431 (deep dive, didn’t know, critical feedback); LC 7724048 (deep dive, missed a commitment, convince someone); GFG 2025 (current role, bug, conflict, negative feedback, learnt something not required — OA-as-R1); GFG Apr 2026 (technical challenge, limited information, conflict — heading “Common Questions”).
What you say
I keep the new prompt on a different product than the story they already heard, using the backup in the table.
They might ask next
If they already heard
On this kind of prompt
Speak this backup
IQVIA
Invent: simplify, challenge, POCs
StratifyLabs: browser inference −30% ML iteration, marketplace of 50+ models
Challenge I solved / significant technical challenge
IQVIA Hybrid RAG
Critical issues
Ylogx Redis + RLS
Beyond your task / exceeded expectations
Ylogx RLS
POCs / multiple solutions
IQVIA Deep Research + Hybrid RAG
Limited information
IQVIA two pages disagree (Dive Deep backup)
Convince someone
IQVIA Backbone
Didn’t know what to do next
IQVIA Learn
One thing you’re not good at
Production Java hours versus intern Python / TypeScript / ROS2 — not a project trophy
Learning new tech
IQVIA LangGraph
Most challenging project
IQVIA Hybrid RAG
Most proud of
One pick: GiftedBooks 3–10 min if the intro used Horizon; otherwise Horizon 17th
10× / scale
Argus 20+ cameras after 89% mAP and 24 FPS
What would you do differently
Stay on the story just told
Keep straight
Ylogx first stories: Ownership, Frugality, Dive Deep. IQVIA first stories: Invent, Learn, Backbone. Argus first stories: Are Right, Highest Standards, Success and Scale. Horizon first stories: Bias for Action, Deliver Results, Earth’s Best Employer.
Limited information uses IQVIA ranking disagreement as the Dive Deep backup. Invent, Learn, and Backbone already cover IQVIA Hybrid RAG, LangGraph, and hybrid-versus-vector.
LC 8362604 extra probes (rate limiting, geofencing) stay on the story they challenged. Independent SDE I Rate Limiter live-round count is 1, not UTA-default leftover coding.
Tighten
Glance at this table, then speak the card. Do not narrate sixteen Leadership Principle names. Do not walk them through other people’s loops.
Unnamed two LP / all 16
They have not named a principle and have not used a leftover stem. Give one deep story, then a second from a different internship. This default pair is a writing lock, not evidence this leftover hour will ask it. There is no list of sixteen stems here.
IE-asked as hour shape, stems unnamed. Nisarg Patel claimed all 16 STAR answers at about 15–20 minutes each and a 15:00–16:00 Pacific hour — those two claims clash; report the clash, do not reconcile it: linkedin.com/posts/nisarg-patel-…. Reddit AUTA Chennai 1qj304z LP-only about 25 min versus 1 hour — this pass could not re-verify (login). LC 7406809 couple LP unnamed “formality” after DSA. LC 6425074 about 40 min STAR unnamed. LC 8014509 two additional LPs unnamed.
What you say — first, Dive Deep
Situation. At Ylogx the SQL RAG chatbot returned SQL and rows, so a demo said the product worked. Every natural-language turn still hit Postgres, and a generated JOIN could cross an organization if isolation lived only in the application.
Task. Find whether slowness was query shape, missing cache, or security in the wrong layer, and make three organizational tiers real on the path the bot already used.
Action. I measured bot database latency instead of rewriting the prompt first. I put RLS and RBAC in Postgres so generated SQL ran as the caller’s database role. I put Redis on that bot path with the organizational tier in the cache key, so a cached answer was still the right tier. Redis down still runs as the caller role, never as table owner.
Result. Bot database latency fell 35 percent. Three-tier isolation sat in the database. Uptime held at 99.9 percent.
What you say — second, Deliver Results
Situation. From February to June 2024 I was Horizon core software for ERC 2024. More than eighty teams. The date does not move. A rover that only sees pixels still collides.
Task. A live operator feed, and a costmap the planner could trust, on the sensors we had.
Action. I owned GStreamer at 60 FPS, ZED 2 mapping at more than two million points per second, and costmap fusion. When the operator feed looked live and occupancy was stale, that split was the ship-blocker. Extra rover hardware was the cut. Skipping occupancy was not.
Result. We placed 17th of more than eighty. Obstacle detection up 40 percent. Collision risk down 55 percent.
They might ask next
If they ask
You answer
Why Redis and RLS together?
Latency and a JOIN leak hid under the same happy demo. Cache without tier leaks. RLS without measuring still leaves every turn on Postgres.
Why fusion, not another sensor?
The date does not wait on a new driver. Extra hardware near the date adds calibration while time alignment is still unsolved.
They want a third
GiftedBooks: student’s own PDF, retrieve-then-generate, doubts from hours to 3–10 minutes.
Intro already used the rover
Second story is Ylogx 40% faster reports with 99.9% uptime — Deliver backup, still Ylogx as a second LP on that intern, not a new intern dump.
They name Ownership, date pressure, and learn (other loops)
Ylogx RLS; Horizon ERC; IQVIA LangGraph. Three projects.
Keep straight
Dive Deep lead is Ylogx Redis and RLS together. IQVIA ranking disagreement is the Dive Deep backup.
17th of 80+ is we. GStreamer, ZED, and costmap software are I.
Formality still gets mechanisms, not a poster of sixteen names. Recruiter “LP-only” is not a reason to drop Java if an editor opens.
Tighten
Stop after the second result unless they ask for a third. Do not mint sixteen stems to fill Nisarg’s claim. Few-deep is the common write-up.
Current role
What you do now. Two FastAPI systems. Not a Leadership Principle dump and not six products.
What they might ask
Current working role and responsibility.
Walk me through your current project architecture and flow.
Situation. I am a software developer intern at IQVIA in Kochi, April 2026 to present. Analysts cannot treat the first web result as truth, and they cannot live inside two-hundred-page BRDs by stuffing a PDF into a prompt.
Task. I own two FastAPI-backed systems. I am an intern, not Head of Platform.
Action. Deep Research uses LangGraph: a planner writes into typed state, tool nodes fan out Firecrawl, Bing, DuckDuckGo, and Playwright, reducers keep surviving documents, and a ranker scores quality, recency, and agreement. The writer synthesizes only from retained evidence, or it abstains. Hybrid RAG is a different path: I ingest the BRD once, query Azure AI Search as hybrid lexical plus vector, let the Azure semantic ranker reorder, and hop a GraphDB for requirement dependencies. GraphDB is the dependency store. LangGraph is the workflow that chooses lexical versus semantic versus hop. LangSmith traces the nodes. A generated test with no overlapping span fails.
Result. Research ranked across 200+ websites. Hybrid RAG covers 200+ page BRDs. Those two “200+” numbers are different families. The resume has no extra IQVIA percentage, so I stop.
They might ask next
If they ask
You answer
What did you implement versus the platform?
I owned graph routing, retrieval mode, and traces. I did not build Azure AI Search.
Why two systems?
Deep Research has no single gold page. Hybrid RAG has a gold long document.
How do you know 200+?
Resume coverage claims. I will not add nDCG or a gold-set size.
What is GraphDB versus LangGraph?
GraphDB stores requirement edges. LangGraph is workflow state.
They already grilled IQVIA in the intro
Do not retell architecture. Lead one failure mode: two high-ranked pages disagree, or hybrid versus vector-only, then stop.
What did you do before this intern?
Ylogx, November 2024 to October 2025: reports, SQL RAG, three-tier RLS. That is the next question, not this answer.
Keep straight
200+ websites is Deep Research coverage. 200+ page BRDs is Hybrid RAG document size.
99.9% and sub-210 ms are Ylogx. IQVIA has no uptime or latency percent on the resume.
LangGraph and LangSmith are IQVIA. The Ylogx chatbot is LangChain SQL RAG.
Internship bullets used Python FastAPI. Live Code is Java. This intern is not a Spring service.
Tighten
Stop after the two systems. Architecture here is the current-role walkthrough. Invent, Learn, and Backbone already cover IQVIA when they name those principles. If they want a previous intern, wait for that question.
Deep-dived a bug
Dive Deep on Ylogx. Two problems under one happy demo: every turn hit Postgres, and a JOIN could cross an organization.
What they might ask
Last time you had to deep dive into a particular bug or task.
A scenario where you had to deep dive to find a problem.
A difficult situation that required you to deep dive.
Situation. At Ylogx the SQL RAG chatbot returned SQL and rows. Two problems hid under that demo. Every natural-language turn hit Postgres, so latency felt like the model thinking. And if isolation lived only in the application, a JOIN through a mapping table could return another organization’s rows while HTTP still said 200.
Task. Find whether slowness was query shape, missing cache, or security in the wrong layer, and make three organizational tiers real on the path the bot already used.
Action. I measured bot database latency instead of blaming the generator. I put RLS and RBAC in Postgres so generated SQL ran as the user’s non-owner database role. I added Redis cache-aside on that bot path. The cache key includes organizational tier and role, so a cross-tier hit is a leak, not a latency win. Redis down still runs as the caller role, never as table owner.
Result. Bot database latency fell 35 percent. Three-tier RLS stayed on. Uptime held at 99.9 percent.
They might ask next
If they ask
You answer
What did you implement?
Measure bot DB; caller role plus RLS; tier and role in the Redis key.
Why Redis, not a bigger RDS?
Repeat natural-language questions and schema lookups are cache-shaped.
Why still RLS after cache?
A cached answer for the wrong org is a leak.
How do you know −35%?
Resume bot-database family. I do not have p95 or TTL on the resume.
What would you add next?
Denied-tier tests and cache-key fixtures from day one.
They already heard Ylogx
IQVIA ranking contradiction: two high-ranked pages disagree; first-hit retrieval is false certainty; agreement plus cite-or-abstain. That is Dive Deep backup.
Keep straight
−35% is Ylogx bot database latency via Redis. sub-210 ms is Ylogx dashboard serving with no LLM on that GET. Different families.
+65%, 40% faster reports, and 30 KPI dashboards are other Ylogx families. This result is −35%, three-tier RLS, and 99.9%.
The Ylogx chatbot is LangChain SQL RAG. LangGraph is IQVIA.
99.9% uptime is Ylogx. IQVIA has no uptime percent on the resume.
Tighten
Stop after −35% and 99.9%. If they already heard Ylogx, switch to IQVIA two pages disagree. Spare third: Horizon — healthy 60 FPS feed with stale occupancy.
Conflict teammate / manager
Technical disagreement on retrieval. Hybrid versus vector-only. No fight scene. Commit on evals.
What they might ask
Last time you had a conflict with a co-worker or manager.
Tell me about a time you had a conflict with a teammate or manager.
A disagreement with your manager.
Conflicting opinions in a team / client with a different point of view.
Situation. At IQVIA the fastest RAG demo on PDFs is a vector index. The BRDs were more than two hundred pages, with clause IDs, tables, and requirement dependencies. Vector-only is a reasonable ship-date instinct. It is the wrong coverage model for a clause ID such as R-141 and for “what breaks if this requirement changes.”
Task. Challenge the retrieval choice through its failure mode, propose a testable alternative, then stop relitigating libraries once the path was chosen.
Action. I stated that similar embeddings are not coverage. Lexical retrieval finds exact IDs. Semantic retrieval finds paraphrased policy. A bounded GraphDB hop finds dependency edges — GraphDB answers structure; Azure AI Search answers citeable prose. I proposed Azure hybrid plus semantic retrieval, GraphDB, and LangGraph choosing among those modes. Once that path was selected, I committed by instrumenting LangSmith: a generated test case with no overlapping source span fails. I did not keep a quiet vector-only side project.
Result. Hybrid RAG processed 200+ page BRDs with traces and generated tests. No extra percentage. Disagree on retrieval. Commit on evals.
They might ask next
If they ask
You answer
What did you implement?
Named the miss (clause IDs / hops), proposed hybrid plus GraphDB, instrumented the fail-if-no-span test.
What did the other person say?
No transcript on the resume. Stay on coverage versus ship date. Vector-only is faster to demo; it misses IDs and hops.
How do you know it worked?
200+ page BRDs plus traces. The resume has no win percentage.
What would you do differently?
Put the eval table in the room before the disagreement and freeze the gold slice earlier.
They already heard IQVIA
Ylogx: RLS in the database versus app-only WHERE. A generated JOIN is legal SQL and illegal for the tenant; the database must refuse the row.
Keep straight
This is Backbone. Technical disagreement. The resume has no manager fight, skip-level, or quoted dialogue.
GraphDB stores requirement dependencies. LangGraph is workflow state.
200+ page BRDs is Hybrid RAG size. 200+ websites is Deep Research coverage.
Tighten
Stop on evals. If they already heard IQVIA, speak Ylogx RLS-in-database versus app-only WHERE. Same-hour with 8430457: this card is disagreement; critical feedback is Argus 73%; exceeded expectations is Ylogx RLS beyond the chatbot ticket.
Negative feedback / mistake
Argus 73% mAP. The evaluation metric was the feedback. 73% is not a number I would page the floor on. The mistake was treating a rising notebook score as launch-ready.
What they might ask
How do you handle feedback, and when was the last time you received negative feedback?
Describe a mistake you made. What did you learn?
Tell me about a time you got feedback / critical feedback / failed.
Situation. Argus is industrial computer vision for PPE compliance. A detector can look promising in a notebook and still be wrong on a floor. At 73% mAP the model misses helmets and also fires junk boxes. Either one trains people to ignore PPE alarms. I treated a rising notebook score as launch-ready. That judgment was the mistake. 73% was not a number I would page the floor on.
Task. Disconfirm the comfortable belief. “Right” meant a held-out mAP I could alert a person on, at a frame rate the live stream could support, with events that could be audited.
Action. I treated 73% as a failed safety-alert gate. I trained and evaluated YOLOv9 on more than 15,000 images and iterated to 89% mAP. I required 24 FPS end-to-end — decode, infer, NMS, I/O — not notebook forward-milliseconds. I wrote detection events to Postgres (camera, class, confidence, time, box), not video blobs. Only after both gates did I containerize more than twenty camera feeds.
Result. 73% to 89% mAP at 24 FPS on 15,000+ images, then 20+ cameras. After alerts existed, site violations fell 50 percent and compliance doubled. Those ops numbers are not mAP. mAP was the gate. −50% / 2× came after the gate.
They might ask next
If they ask
You answer
What did you implement?
Data plus eval gate, then FPS, then cameras. Events in Postgres, not video blobs.
Why was 73% unshippable?
Missed helmets plus junk boxes both train people to ignore alarms. Scaling 73% multiplies ignored alarms.
Who gave the feedback?
The evaluation metric. I do not attach that number to a person. No skip-level scene on the resume.
How do −50% and 2× relate to mAP?
They are site outcomes after alerts existed. mAP is held-out detector quality.
What would you add next?
Hold out a camera earlier. Future work, not a shipped claim.
They already heard Argus
Ylogx measured bot-DB latency versus the anecdote “the model is slow” (−35% after Redis, still with RLS).
Keep straight
73% mAP is Argus. It is not shippable for floor alerts. The gate is 89% mAP and 24 FPS, then 20+ cameras.
Argus detector budget is 24 FPS. Horizon camera pipeline is 60 FPS.
−50% violations and 2× compliance are we, after alerts existed. mAP and 24 FPS are I.
The fail is 73% mAP. The resume has no missed ERC and no missed 99.9%.
Tighten
Stop after the gate and the ops numbers, labeled as after alerts. If they want Standards next: same Argus gate. If they want Scale: 20+ cameras only after 89% and 24 FPS. Same-hour 6806195: after this Argus card, critical issues is Ylogx Redis+RLS; beyond-task is Horizon GStreamer + costmap.
Learned something not required
IQVIA LangGraph. Curiosity needed a stop: an observable pass or fail. The resume has no hour count for this learning.
What they might ask
The last time you learnt something that wasn’t required at your job — what was your way of learning, and how much time did it take?
IE-askedResume-derived GFG 2025 BR full sentence: GFG 2025 (OA-as-R1). Adjacent Learn family: LC 7035289 “How would you learn a new Technology” (separate card). LC 7623949 named Learn & Be Curious (LP name, not a STAR sentence). Rudraksh “quickly learn something new” is their earlier live — not leftover-hour default.
What you say
Situation. At IQVIA the work needed multi-agent research, hybrid retrieval, requirement dependencies, and a stop condition you can see. Calling a model from an API was not the job. Tutorial-complete is not architecture.
Task. Learn enough to rank 200+ websites and retrieve evidence from 200+ page BRDs without inventing citations. Curiosity needed a stop: an observable pass or fail.
Action. I learned LangGraph typed state and checkpoints so one failed Playwright branch does not erase Bing. I drew planner, parallel tools, ranker, writer, and failure edges before adding another tool. I learned adaptive retrieval on the BRD path: lexical for clause IDs, semantic for policy language, GraphDB hops for dependencies. I used LangSmith traces as the stop — empty scrape reaching the writer is visible; a generated test with no overlapping span fails.
Result. 200+ sites ranked. 200+ page BRDs processed with traces and test-case generation. No extra IQVIA percentage. I do not add an hour count; the resume does not have one.
Draw the graph, instrument one fail, then add a tool. Not a certificate.
How much time?
The resume has no hour count. I will not add one.
How do you know you learned it?
Traces show an empty scrape; the writer does not get to be fluent anyway.
What would you add next?
Freeze gold questions with expected citations before a fourth search tool.
They already heard IQVIA LangGraph
Horizon ROS2 + GStreamer + ZED 2 — field lighting and timing versus Gazebo.
Keep straight
LangGraph and LangSmith are IQVIA. The Ylogx chatbot is LangChain SQL RAG.
200+ sites and 200+ page BRDs are two families. IQVIA has no extra percent.
This is Learn. Invent is Hybrid RAG coverage. Backbone is hybrid versus vector.
Tighten
Stop on traces. If Invent, Learn, and Backbone have already spent IQVIA in this hour, speak Horizon ROS2 + GStreamer + ZED 2.
Missing a deadline
Horizon sequencing. The ERC date does not move. The date was met. The resume has no missed ERC and no missed 99.9%. Speak early, specific risk — not an invented miss.
What they might ask
Missing a deadline.
A time you missed a commitment or deadline.
IE-askedResume-derivedLC 8362604 AUTA BR topic list. LC 7724048 BR: missed a commitment or deadline. IE.in 2024-grad BR list: IE.in 2024-grad. GFG 2022 FTE “failed a deadline” is year-out — mapping only, not this leftover hour. LC 7563011 “how will you communicate if you think you will miss deadline” is their R2, earlier live. Resume: ERC 17th / 80+.
What you say
Situation. ERC 2024 does not move. More than eighty teams. I was Horizon core software from February to June 2024. A rover that only sees pixels still collides. A pretty Gazebo run is not the contest.
Task. Sequence the work that had to exist on the day: a live operator feed, and a costmap the planner could trust. Flag risk early and specifically. Extra hardware near the date adds drivers and calibration while time alignment is still unsolved.
Action. I shipped GStreamer at 60 FPS so the operator saw the newest frame, not a stalled historical stream. I moved ZED 2 mapping at more than two million points per second into the rover frame with timestamps. Fusion wrote occupied / free / uncertain costmap cells. When occupancy lagged a healthy feed, I named that split out loud: the costmap is stale; fusion is the path; extra hardware is the cut. We downsampled for visualization and kept local planning density.
Result. We placed 17th of more than eighty. Obstacle detection up 40 percent. Collision risk down 55 percent. The date was met.
They might ask next
If they ask
You answer
What did you implement?
Sequence feed and costmap; name stale occupancy; cut extra hardware.
Did you miss the date?
No. The date was met. 17th of 80+.
How would you communicate a late ship-blocker?
Early and specific: “the feed is late,” or “the costmap is stale.”
What is I versus we?
I owned camera / mapping / costmap software. Placement is the team.
What would you add next?
Name the ship-blocker in week one and time-box sensor versus fusion.
They already heard the rover
Ylogx 40% faster reports and 99.9% — flag “RLS is not in the database yet,” still without an SLA miss.
Keep straight
Horizon camera pipeline is 60 FPS. Argus detector budget is 24 FPS.
+40% obstacle and −55% collision are Horizon rover software. They are not Argus.
17th / 80+ is we. The resume has no missed ERC and no missed 99.9%.
Ownership first story is Ylogx RLS. Horizon GStreamer + costmap is Ownership backup.
Tighten
Stop after 17th and the software numbers. Same-hour 8362604: this is Deliver; challenge stays IQVIA; identified stays Ylogx JOIN-leak.
Problem I identified
Ylogx JOIN-leak. The intern ticket could have ended when the chatbot returned rows. Isolation on the shared path was the product.
What they might ask
A problem I identified myself.
IE-askedResume-derivedLC 8362604 AUTA BR topic list (topic, not a quoted interviewer sentence). Adjacent “issue not your task” family is the beyond-task card.
What you say
Situation. At Ylogx my intern ticket could have ended when the chatbot returned rows and the report builder drew a table. I looked at the shared path those products already used. A generated JOIN can be syntactically valid and still expose another organization’s facts through a mapping table. App-only WHERE dies on the first missed handler or skipped helper.
Task. Make three organizational tiers real where SQL actually ran, not only where the React UI hid a column.
Action. I put RLS and RBAC in Postgres so generated SQL executed as the caller’s database role, not a table owner. A JOIN cannot outrun the policy. The nightmare is HTTP 200 with another org’s rows. The same identity mapping served reports and the bot.
Result. Three-tier RLS on the shared path. Uptime held at 99.9 percent.
They might ask next
If they ask
You answer
What did you implement?
Named the JOIN-leak class; shipped role plus RLS on the shared path.
Why not app-only WHERE?
Copied handler, generated SQL, mapping-table JOIN. The database has to refuse the row.
How do you know?
Resume three-tier RLS and 99.9%. No invented incident count.
Where does Redis −35% sit?
Dive Deep / Frugality. Not a second Ownership trophy on this card.
They already heard Ylogx RLS
Horizon: a healthy 60 FPS camera with stale occupancy. I identified that the live video was not the planner’s map.
Keep straight
This is Ownership. Dive Deep on the same intern is Redis and RLS together as the bug card — latency plus leak under one demo.
99.9% uptime is Ylogx. IQVIA has no uptime percent.
40% faster reports is a Deliver-family Ylogx number. This card’s result is three-tier RLS and 99.9%.
Tighten
Stop after three-tier and 99.9%. Same-hour 8362604: this is Ownership; challenge stays IQVIA; deadline stays Horizon.
Simplifying a process
IQVIA Hybrid RAG. Stuffing a 200+ page BRD looks simple and produces confident citations to the wrong annex. Simplify the untraceable work: retrieve a small evidence set, cite or abstain.
What they might ask
Simplifying a process.
A time you improved an internal process (year-out wording elsewhere; speak the same Hybrid RAG story).
IE-askedResume-derivedLC 8362604 AUTA BR. GFG 2022 FTE “improved an internal process” is year-out, different wording — mapping only.
What you say
Situation. At IQVIA, analysts could not live inside 200+ page BRDs. Stuffing the PDF into a prompt looks simple and produces confident citations to the wrong annex. Hand-writing tests from those BRDs does not scale and still cannot tell whether the ranker, an empty retrieve, or the writer failed.
Task. Simplify the untraceable work: retrieve a small evidence set, cite or abstain, and generate tests a human can audit. Deleting steps without a check is not simplify.
Action. I ingested each BRD once, chunked on headings, requirement IDs, and table boundaries. One hybrid Azure AI Search query covers lexical IDs and vectors. Semantic ranker reorders for meaning. A bounded GraphDB hop handles “depends on.” LangGraph picks the mode. The writer sees a small evidence set, not 200 pages. LangSmith gold items carry expected citation IDs; test-case generation is itself a traced job. No overlapping span fails.
Result. 200+ page BRDs processed with traces and generated tests. No extra retrieval percentage.
They might ask next
If they ask
You answer
What did you implement?
Chunk-once, hybrid plus semantic, GraphDB hop, cite-or-abstain, traced tests.
Why is stuffing not the simple path?
Tokens, invented citations, no way to see which node failed.
How would this grow?
Do not re-embed a BRD on every question; bound hops. I did not shard Azure or the graph.
What would you add next?
Freeze the citation gold slice before a fourth tool.
They already heard IQVIA Invent
StratifyLabs: browser inference cut ML iteration 30 percent; marketplace of 50+ models so a notebook demo is not the product. −30% is iteration time, not FPS.
Keep straight
GraphDB stores requirement edges. LangGraph is workflow state.
StratifyLabs −30% is ML iteration time in the browser. It is not Argus 24 FPS and not Horizon 60 FPS.
Ylogx Redis is Frugality / Dive Deep. It is not this Invent story.
Tighten
Stop after 200+ page BRDs and traces. Same-hour 8362604: if challenge already used IQVIA Invent, switch this card to StratifyLabs.
Challenge I solved
IQVIA Hybrid RAG. The challenge was coverage, not a prettier paragraph. Same first story if they say “significant technical challenge.”
What they might ask
A challenge I solved.
Tell me about a time you faced a significant technical challenge.
A technically complex discussion / challenging implementation.
Was that actually challenging? Rate limiting? Geofencing? Trade-offs?
IE-askedResume-derivedLC 8362604 AUTA BR topic list. GFG Apr 2026 HM: GFG Apr 2026, heading “Common Questions.” Adjacent: IE.in 2024-grad “challenging implementation”: IE.in 2024-grad. LC 7035289 “Most Challenging Project” is a separate card. LC 8362604 extra probes (rate limiting, geofencing) are follow-ups on the story they challenged, not a Rate Limiter LLD.
What you say
Situation. IQVIA Hybrid RAG on confidential 200+ page BRDs. The challenge was coverage, not a prettier paragraph. Vector-only cosine misses exact clause IDs. Keyword-only misses paraphrased policy. Stuffing invents citations. “What breaks if this requirement changes” is a dependency edge, not a nearest neighbor.
Task. Make retrieval fail visibly when evidence is missing, and make the right retrieve mode available when the question is an ID, a policy paraphrase, or a hop.
Action. I shipped Azure AI Search hybrid plus semantic, a GraphDB for one-or-two-hop dependencies with a visited-set, and LangGraph choosing lexical versus semantic versus hop. Writer sees a small set. Cite or abstain. LangSmith: generated tests fail without an overlapping span. That span-less fail is the challenge, not “the model sounded good.”
Result. 200+ page BRDs on that path. No extra percentage.
They might ask next
If they ask
You answer
What did you implement?
Hybrid plus semantic plus GraphDB plus span gate.
Was that actually challenging?
Stay here: clause-ID miss, span-less test, two pages that disagree. Same product.
Rate limiting / geofencing?
Stay on IQVIA clause IDs and span overlap. Those probes are on this story. Independent SDE I Rate Limiter live-round count is 1, not UTA-default leftover coding.
Why not increase vector top-k?
Neighbors still miss IDs and hops.
How would this grow?
Small evidence sets; bound hops. I did not shard Azure or the graph.
They already heard IQVIA Invent
Horizon: more than two million points per second into a live costmap that must not go stale. Deliver physics.
Keep straight
Skills line says Graph DB; the intern bullet says GraphDB. I speak GraphDB. I do not fake a named cluster.
200+ page BRDs is this challenge. 200+ websites is Deep Research coverage — a sibling system, not this metric.
FAISS is skills-listed. IQVIA intern retrieval is Azure AI Search Hybrid + Semantic + GraphDB.
Tighten
If they challenge “was that actually challenging,” stay on clause IDs and span-less tests. Same-hour 8362604: simplifying goes to StratifyLabs, not a second IQVIA Invent trophy. IE.in 2024-grad: complex discussion and challenging implementation are one IQVIA Invent story, not two.
Critical issues
Ylogx Redis and RLS together. The critical issue was two-sided: every turn still hit Postgres, and a JOIN could cross an organization.
What they might ask
Tell me about a time you faced critical issues.
IE-askedResume-derivedLC 6806195 AUTA BR as third. Same hour also asked feedback (Argus card) and beyond-task (next card). Extra probes on all unnamed — stay on this story.
What you say
Situation. Ylogx SQL RAG looked healthy: the bot returned SQL and rows. The critical issue was two-sided. Every natural-language turn still hit Postgres. And a JOIN could cross an organization if isolation was only in the application. A confident answer could be slow, wrong-grained, or leaked.
Task. Treat latency and isolation as one path, because a cache without a tier, or a role without RLS, each makes the other worse.
Action. I measured bot database latency. I put three-tier RLS and RBAC in Postgres so generated SQL ran as the caller role. I put Redis on the bot path with tier in the key. I kept showing the generated SQL so wrong grain was visible even when tenant isolation was correct.
Result. Bot database latency fell 35 percent. Three-tier isolation sat in the database. Uptime held at 99.9 percent.
They might ask next
If they ask
You answer
Why both Redis and RLS?
Cache without tier leaks. RLS without measuring still leaves every turn on Postgres.
What happens if you scale without RLS?
More traffic on a leak is still a leak.
What would you add next?
Deny fixtures and invalidation next to RLS from day one.
Dashboard sub-210 ms?
Different Ylogx family: dashboard GET with no LLM. This result is bot-DB −35%.
They already heard Ylogx Dive Deep
IQVIA ranking contradiction — Dive Deep backup.
Keep straight
This is Dive Deep. Ownership is RLS as the product beyond the chatbot ticket. Frugality is Redis versus a bigger RDS. Same intern, different prompts — if they already used Ylogx here, beyond-task in the 6806195 hour is Horizon.
−35% is bot database latency. sub-210 ms is dashboards. 99.9% is uptime. Three families.
Tighten
Stop after −35% and 99.9%. Same-hour 6806195: feedback is Argus 73%; beyond-task is Horizon GStreamer + costmap so Ylogx is not used twice.
Beyond your task
Ylogx RLS. The visible intern ticket was chatbot answers and custom reports. Isolation on the shared path was beyond that ticket. If they already used Ylogx, speak Horizon GStreamer + costmap.
What they might ask
Went beyond your task or job.
Exceeded expectations.
Going above and beyond responsibilities.
IE-askedResume-derivedLC 6806195; LC 8430457 exceeded expectations. Bhavya LinkedIn paraphrase: linkedin.com/posts/bhavya-…. GFG sde-1-17 “extra for customers” maps to GiftedBooks Customer — pick one if they want customer-shaped, do not mash it into this Ownership card.
What you say
Situation. At Ylogx the visible intern ticket was chatbot answers and custom reports. The demo returned rows. That was not the product risk. Operators across three organizational tiers would share warehouse facts. Application-only checks would let a generated JOIN cross a tier.
Task. Own isolation on the path the bot and the reports already shared, even though nobody wrote “Postgres RLS” on the chatbot ticket.
Action. I put RLS and RBAC in Postgres so generated SQL ran as the user’s database role. NestJS still did route-level RBAC; the database remained the last gate. FastAPI stayed on SQL RAG only. Dashboard GETs never called the model. Same identity mapping for reports and the bot.
Result. Uptime held at 99.9 percent. Three-tier RLS on the shared path.
They might ask next
If they ask
You answer
What was the ticket versus the product?
Ticket = rows in a demo. Product = caller role plus RLS.
What did you not own?
A named SRE on-call title. Dashboard GETs never called the model; that was already the serving path.
What would you add next?
Deny-tier tests from day one.
They already used Ylogx (6806195 critical issues)
Horizon Ownership backup: GStreamer 60 FPS and ZED-to-costmap were beyond a screenshot camera demo and inside the ERC result.
They want customer-shaped extra
GiftedBooks RAG + PYQ beyond a pretty VR lab — hours to 3–10 min. Pick that or this Ylogx card, not both mashed.
Keep straight
This is Ownership. Redis −35%, SQL RAG +65%, 30 KPIs, and sub-210 ms are other Ylogx families — not this result.
Horizon GStreamer + costmap is the Ownership backup, not the Ownership first story.
99.9% is Ylogx. GiftedBooks uptime is 99.5%.
Tighten
Standalone beyond-task = this Ylogx card. In the 6806195 hour, if critical issues already used Ylogx Dive Deep, switch to Horizon GStreamer + costmap.
POCs / multiple solutions
IQVIA. A short series of proofs: which retrieve and which orchestration actually fail closed. Intern: I owned the POCs and the eval stop, not a research org.
What they might ask
Can you describe a complex problem you encountered that required in-depth research, development of proof of concepts, and exploration of multiple solutions?
IE-askedResume-derivedLC 7563011 AUTA Bengaluru HM as third (no BR named then), full sentence. Work about 25 min; hour about 35 min; no DSA named. The candidate tagged LPs on that post; the lock for this sentence is Invent. Dhananjai “complex problem you faced” is adjacent, not this sentence: Dhananjai Medium.
What you say
Situation. IQVIA needed ranked research across many site families and grounded answers on 200+ page BRDs. One happy-path demo was not enough. Each wrong architecture failed in a different way.
Task. Run a short series of proofs: which retrieve and which orchestration actually fail closed.
Action. One scraper failed across site families — Firecrawl, Bing, DuckDuckGo, and Playwright die on different pages, so the Deep Research graph fans out in parallel with checkpointed state; a dead Playwright branch does not erase Bing. Vector-only missed clause IDs — the BRD POC became Azure hybrid plus semantic plus GraphDB hops. A linear chain could not retry or choose a hop. Stuffing the BRD failed on tokens and citation honesty. LangSmith made the stop visible: a generated test without an overlapping span fails. LangGraph is the workflow. GraphDB is the requirement-dependency store. Those are two different graphs.
Result. 200+ websites ranked. 200+ page BRDs with traces and generated tests. No extra percentage.
They might ask next
If they ask
You answer
What did you discard?
One scraper; vector-only; linear chain; stuffing the BRD.
What did you implement?
Parallel tools plus hybrid plus GraphDB plus span-fail.
What would you add next?
Freeze an eval slice before the fourth tool; require citations in the writer type earlier.
How would this grow?
Cache scrapes; fewer full Playwright renders. I did not shard.
They already heard IQVIA Invent
StratifyLabs browser inference −30% iteration and marketplace 50+.
Keep straight
LangGraph is the workflow. GraphDB is the requirement-dependency store.
IQVIA has no latency percent, nDCG, or gold-set size on the resume.
This hour shape (HM, no DSA named on 7563011) is other-loop evidence, not a forecast that your leftover hour is an HM with no coding.
Tighten
Deep Research tools are named as POCs here. Do not retell the whole Learn STAR (planner philosophy) unless they ask Learn next. If Invent, Learn, and Backbone already used IQVIA, speak StratifyLabs.
Limited information
IQVIA: two high-ranked pages disagree. First-hit retrieval looks like certainty. This is the Dive Deep backup on IQVIA retrieval — not Invent, Learn, or Backbone again. If those three already appeared, speak Horizon field versus Gazebo.
What they might ask
Describe a situation where you had to solve a problem with limited information.
IE-askedResume-derived — Common Questions caveat. GFG Apr 2026 HM after two tech lives: geeksforgeeks.org/…/sde-1-amazon-experience-verdict-selected. Page heading is “Common Questions” under Round 3, not first-person “I was asked.” This exact sentence was not found in other opened leftover-hour bodies.
What you say
Situation. On IQVIA Deep Research there is no single gold page. Two high-ranked sources can both look credible and disagree. First-hit retrieval then looks like certainty. An empty scrape can still feed a fluent writer. That is limited information, not a missing library.
Task. Rank in a way that disagreement is visible, and refuse to average two policies into vague prose.
Action. Parallel tools bring independent indexes. The ranker scores source quality, recency, and agreement, and treats syndicated mirrors as one vote, not extra confirmation. The writer synthesizes only from retained evidence. If spans conflict or no span remains, the system abstains rather than smoothing the contradiction. LangSmith shows whether an empty body reached the writer.
Result. Research ranked across 200+ websites. I do not add an accuracy percentage. When sources disagree, agreement-or-abstain is the product.
Surface disagreement to the analyst instead of only abstaining — would-do, not a shipped UI claim unless they ask.
Is this the hybrid-versus-vector story?
No. That is Backbone. This card is ranking when two pages disagree.
IQVIA Invent, Learn, and Backbone already appeared
Horizon field versus Gazebo: lighting, dust, and timing are not in the sim; the limited information is whether occupancy is current.
Keep straight
IQVIA first stories are Invent, Learn, and Backbone. Ranking disagreement is the Dive Deep backup on IQVIA retrieval.
Mirrors count as one vote, not extra confirmation.
Ylogx RLS is a different intern. This card does not import authorization theatre onto IQVIA.
GFG Apr 2026 heading is “Common Questions.” That is not the same as “I was asked in my loop.”
Tighten
Stop on cite-or-abstain. If IQVIA already appeared three times, speak Horizon field versus Gazebo. Backup conflict remains hybrid versus vector on a different prompt.
Convince someone
IQVIA Backbone. Conflict is the disagreement. This card is the buy-in, then commit. Same locked story: hybrid versus vector, then evals.
What they might ask
A scenario where you had to convince someone of your approach.
Situation. IQVIA stakeholders could ship a vector-only RAG demo faster. I needed them to accept hybrid retrieval plus GraphDB hops for 200+ page BRDs. I was an intern. I do not win by volume.
Task. Convince on a failure mode they can reproduce, then commit wholly once the path is chosen.
Action. I showed the miss: a clause ID such as R-141 is a lexical problem; “retention policy” is a semantic problem; “what depends on this requirement” is a GraphDB hop. I proposed Azure hybrid plus semantic plus a bounded hop, with LangSmith as the decision record: a generated test with no overlapping span fails. After the path was selected, I stopped the library debate and instrumented that path.
Result. Hybrid RAG on 200+ page BRDs with traces and generated tests. Convince on coverage. Commit on evals.
They might ask next
If they ask
You answer
What did you implement?
Failure-mode table, then evals as the contract.
What is I versus we?
I argued coverage and instrumented. I did not overrule a named manager.
What would you add next?
Put the eval table in the room first.
They already heard IQVIA
Ylogx RLS in the database versus app-only WHERE — convince that the database is the last gate because generated SQL will not remember org_id.
Keep straight
No skip-level win and no invented quote. Commit means the vector-only prototype stops.
Same-hour 7724048: deep dive is Ylogx Redis+RLS; missed commitment is Horizon sequencing.
8014509 two unnamed extra LPs: default pair remainder — Deliver Horizon, or GiftedBooks 3–10 min. Those two stems stay unnamed.
Tighten
Stop on commit. If they already heard IQVIA Backbone, speak Ylogx RLS-in-DB versus app-only WHERE.
Didn’t know what to do next
IQVIA Learn. List options, pick with a stop condition, learn the unknown from traces rather than from another tutorial.
What they might ask
Tell me a time when you didn’t know what to do next or how to solve a challenging problem. How did you learn what you don’t know? What were the options you considered? What was the outcome?
IE-askedResume-derivedLC 7850431 their R4 Behavioural (heading is Behavioural — do not upgrade it to Bar Raiser). Same hour: deep dive (Ylogx) and critical feedback (Argus).
What you say
Situation. At IQVIA I did not initially know the next node for multi-agent research and long-document retrieval. Several architectures all looked plausible on a whiteboard.
Task. List options, pick with a stop condition, learn what I did not know from traces rather than from another tutorial.
Action. Options I considered: one scraper versus parallel Firecrawl, Bing, DuckDuckGo, and Playwright; a linear chain versus LangGraph with checkpointed state; vector-only versus Azure hybrid plus GraphDB; stuffing the BRD versus retrieve-and-cite. I learned the unknowns by drawing planner / tools / ranker / writer and by failing closed: empty scrape must not become a fact; a generated test with no overlapping span fails. LangSmith traces showed which node was wrong.
Result. 200+ sites ranked. 200+ page BRDs with traces. The outcome of not-knowing was a stop condition, not a larger prompt.
They might ask next
If they ask
You answer
What options?
Four: one scraper versus parallel tools; linear chain versus LangGraph; vector-only versus hybrid plus GraphDB; stuffing versus retrieve-and-cite.
How did you learn the unknown?
Trace the empty scrape. The writer does not get to be fluent anyway.
When blocked, who did you ask?
Traces and evals first, then the teammate who owns the adjacent node. No invented named mentor.
What would you add next?
Freeze gold questions before the fourth tool.
They already heard IQVIA Learn
Horizon field versus Gazebo / ROS2 stack — you do not know if occupancy is current until the field run.
Keep straight
LangGraph is IQVIA. The Ylogx chatbot is LangChain SQL RAG.
Same-hour 7850431: deep dive is Ylogx; feedback is Argus 73%. Three products.
Tighten
Stop on the stop condition. If Invent, Learn, and Backbone already used IQVIA, speak Horizon field versus Gazebo.
One thing you’re not good at
Intern-level honesty. Production Java hours are below intern Python, TypeScript, and ROS2 hours. Live Code is Java. Internships are not Spring. Close the gap in the editor. This is not a fake “I work too hard,” and it is not a project trophy.
What they might ask
Tell me something that you are not good at.
IE-askedResume-derivedLC 7035289 AUTA BR as third. Adjacent LP: Earn Trust (vocally self-critical). Not a Ylogx / IQVIA / Argus / Horizon first story.
What you say
Situation. Live Code in this process is Java. My internships and shipped projects were Python FastAPI, LangGraph, YOLOv9, TypeScript NestJS and React, and ROS2. I have fewer production Java hours than those stacks. That gap is real.
Task. Name it without converting it into a fake strength, and close it in the editor rather than by claiming a JVM internship I did not have.
Action. I write the solution in Java — collections, heap, maps, careful long where sums overflow. Internship bullets used Python FastAPI, TypeScript NestJS, and ROS2. Those internships are not Spring or Kafka services.
Result. The honest intern-level answer is the hours split. I close it by producing working Java in the room, not by a story that I work too hard.
They might ask next
If they ask
You answer
What are you doing about it?
Java in the editor this hour.
Did Ylogx or IQVIA run on the JVM?
No. Those internships shipped Python and TypeScript. Live Code is Java.
They want a product shortcoming instead
Argus 73% as a ship bar — only if they did not already hear the mistake card.
They open GiftedBooks GitHub
README currently describes AegisAI. Resume wins. That disclosure is Earn Trust, not this weakness unless they open GitHub.
Keep straight
Internship bullets used Python FastAPI, TypeScript NestJS, and ROS2. Live Code is Java. Spring, Kafka, and EKS are not on those bullets, so do not describe the internships as if they were.
This card is not Ownership, Invent, Dive Deep, or Are Right. Same-hour 7035289: Learn = IQVIA LangGraph; most challenging in that hour = Horizon ERC so IQVIA is not Learn and Invent back-to-back.
Tighten
Name the hours split, then stop. If they open an editor, write Java. No “I work too hard.” No fake SDM weakness. No invented manager who said you are bad at X.
Learning new tech
IQVIA LangGraph as the example of how you learn: draw the failure edges, instrument one fail, then add a tool. Tutorial-complete is not the stop.
What they might ask
How would you learn a new technology?
IE-askedResume-derivedLC 7035289 AUTA BR. GFG 2025 learnt-not-required is the sibling card (past tense, intern example). Rudraksh “quickly learn something new” is their R2 — earlier live, not leftover default.
What you say
Situation. A new stack is not a completed tutorial. At IQVIA the new pieces were LangGraph state, Azure hybrid retrieval, GraphDB hops, and LangSmith as a stop. Calling a model from an API was not the architecture.
Task. Learn by drawing the failure edges, then instrument one fail, then add a tool. Tutorial-complete is not the stop.
Action. I drew planner, parallel tools, ranker, writer. I learned checkpoints so a dead branch does not wipe the run. I learned which retrieve mode matches which question: lexical IDs, semantic policy, GraphDB dependencies. I used traces: empty scrape must not become a sentence; no overlapping span fails the generated test. That is how I would learn the next shop technology too — smallest mechanism that can fail closed, then grow.
Result. 200+ sites and 200+ page BRDs with traces. No invented hour count. No certificate.
They might ask next
If they ask
You answer
Way of learning?
Graph on paper, one fail, then a tool.
How much time?
The resume has no hour count.
Is Java the new technology?
Java for Live Code is the hours gap on the not-good-at card. This card is how I learned LangGraph / hybrid retrieval with a stop.
They already heard IQVIA Learn
Horizon ROS2 + GStreamer + ZED 2 — learn by field failure (stale occupancy, lighting), not by another Gazebo screenshot.
Keep straight
LangGraph is IQVIA. The Ylogx chatbot is LangChain SQL RAG.
This is Learn. The not-good-at card is Java hours. They are different prompts.
Tighten
Stop on traces. If IQVIA Learn already appeared, speak Horizon ROS2 + GStreamer + ZED 2.
Most challenging project
Pick one. First story: IQVIA Hybrid RAG. If they already heard IQVIA, speak Horizon ERC. Do not list three.
Situation. The most challenging project I will defend is IQVIA Hybrid RAG on 200+ page confidential BRDs. Ranking research across 200+ sites is the sibling system; the hardness here is mixed evidence in one long document plus the cost of a fluent wrong citation in a regulated setting.
Task. Coverage you can audit: IDs, paraphrases, and dependencies, with a fail if the span is missing.
Action. Ingest once. Azure hybrid plus semantic. Bounded GraphDB hops. LangGraph selects the mode. Cite or abstain. LangSmith tests fail without overlap. Secrets stay out of traces. That combination is the challenge — not a larger context window.
Result. 200+ page BRDs with traces and generated tests. No extra percentage.
What you say — if they already heard IQVIA
Situation. ERC 2024 does not move. More than eighty teams. I was Horizon core software. A healthy camera feed with a stale costmap still collides.
Task. Live operator video and a planner map that matched the world, on the sensors we had, by contest day.
Action. GStreamer at 60 FPS. ZED 2 mapping at more than two million points per second. Costmap fusion. Extra hardware was the cut.
Result. We placed 17th of more than eighty. Obstacle detection up 40 percent. Collision risk down 55 percent.
They might ask next
If they ask
You answer
Why this one?
Long confidential BRDs plus ranking and citation cost. Stay on Hybrid RAG unless they already heard IQVIA.
What did you implement?
Hybrid plus GraphDB plus span gate. Secrets out of traces.
What did you not own?
Azure AI Search itself. No IQVIA latency or uptime percent on the resume.
Same-hour 7035289 after Learn already used IQVIA
Prefer Horizon for this pick so Learn and Invent are not back-to-back IQVIA.
Keep straight
One project. IQVIA Hybrid RAG first. Horizon ERC if IQVIA already appeared.
Ylogx year of production BI is Ownership when they ask isolation, not a steal of Invent.
StratifyLabs is Think Big (marketplace 50+). Invent first story is IQVIA; StratifyLabs is Invent backup.
Tighten
Pick one and stop. Do not recite every intern in one breath.
Most proud of
Pick one. If the intro already used Horizon, speak GiftedBooks hours → 3–10 min. If they have not heard the rover, speak Horizon 17th / 80+. Do not list three.
What they might ask
What are you most proud of?
What’s a task you’re most proud of?
Resume-derived — stem not found as leftover-hour interviewer wording in the merge. Freeze: Ruchi FTC “What’s a task you’re most proud of?” was their R2 — earlier live. Speakable from the Aug 2026 PDF anyway. Default leftover-hour pick if the rover already appeared in an intro: GiftedBooks. If they have not heard the rover: Horizon 17th / 80+.
What you say — if the intro already used Horizon
Situation. GiftedBooks is a VR learning suite. The customer is a student who already has PDFs and still loses hours on doubts. A lab that cannot answer from their upload is decoration. A generic tutor that sounds like a textbook is the wrong customer.
Task. Cut time-to-answer without inventing content, and keep the API fast enough to use while studying.
Action. I shipped RAG over the student’s PDF: chunk and embed once, retrieve a small user-scoped top-k, generate from a supporting span or abstain. I kept generation small for a sub-300 millisecond API. PYQ topic ranking is a separate deterministic pass over past papers — not the generator guessing the syllabus. Per-user isolation so another student’s notes never retrieve.
Result. Doubts from hours to 3–10 minutes. Reading efficiency +35%, engagement +50%, comprehension 2.5×. sub-300 ms API. 99.5% uptime. I owned RAG and PYQ, not every 3D asset.
What you say — if they have not heard the rover
Situation. From February to June 2024 I was Horizon core software for ERC 2024. More than eighty teams. The date does not move.
Task. A live operator feed and a costmap the planner could trust.
Action. GStreamer at 60 FPS. ZED 2 mapping at more than two million points per second. Costmap fusion on the sensors we had.
Result. We placed 17th of more than eighty. Obstacle detection up 40 percent. Collision risk down 55 percent.
They might ask next
If they ask
You answer
What did you implement? (GiftedBooks)
RAG plus abstain plus PYQ analytics. Not every avatar asset.
Why RAG, not dump the PDF?
Tokens, invented citations, and it blows sub-300 ms.
How do you know 3–10 min?
Resume claim. I will not invent student n or a study protocol that is not on the PDF.
They open GitHub on GiftedBooks
README currently describes AegisAI. Resume wins. One sentence, then stay on the PDF metrics. That disclosure is Earn Trust.
Both Horizon and GiftedBooks already used
Spare only: Argus 73→89 then 20+ cameras. Still one project, not a list of three.
Keep straight
GiftedBooks uptime is 99.5%. Ylogx uptime is 99.9%. Different products.
GiftedBooks API is sub-300 ms. Ylogx dashboards are sub-210 ms. Different APIs.
LangSmith is IQVIA. GiftedBooks RAG does not inherit LangSmith.
GiftedBooks is Customer Obsession / Earn Trust, not Invent or Think Big.
Ruchi FTC R2 is earlier live, not this leftover hour.
Tighten
One proud project. GiftedBooks if the rover already appeared; Horizon 17th if it has not. Do not list three.
10× / scale
If they ask scale: Argus 20+ cameras after 89% mAP and 24 FPS. Honest next step is more workers or batching if latency holds. It is not 73% scaled, and it is not 10× FPS. Interviewer-quoted “10×” in opened leftover-hour bodies: I could not verify.
What they might ask
How would this scale? / How did you scale it?
What if traffic were 10×?
Resume-derived — interviewer wording I could not verify (merge leftover-hour bodies n = 0). LC 8362604 candidate self-note “stories didn’t have massive scale” is not a question. igreaper “How did you scale it?” is their R2 — earlier live: igreaper.medium.com/…. If they still ask scale, resume lock: Argus 20+ cameras after 89% mAP and 24 FPS — Success and Scale.
What you say
Situation. One-camera detection does not create industrial safety. At more than twenty cameras, wrong alerts scale harm as well as benefit. A detector at 73% mAP would multiply misses and junk boxes until people learned to ignore the system.
Task. Make the detector ready before multiplying it. Preserve 24 FPS. Leave an audit trail. I did not shard Postgres.
Action. I did not scale at 73%. I moved YOLOv9 to 89% mAP at 24 FPS on more than 15,000 images, kept NMS, and only then containerized more than twenty feeds. Workers drop stale frames rather than queue delayed alerts. Postgres stores events — camera, class, confidence, time, box — not JPEG blobs. Honest 10× from here is more workers or batching if latency holds, still behind the same gates. It is not 10× FPS, and it is not 73% on 200 cameras.
Result. After the gate: 20+ cameras. After alerts existed: −50% violations, 2× compliance. Those ops numbers are not mAP. Quality, then scale.
They might ask next
If they ask
You answer
What did you implement?
Gate, then containers, events not video. Drop stale frames.
Did you shard?
No. First: eval gate, FPS, workers, events.
What would you add next?
Per-camera error budget so a noisy camera cannot train the floor to ignore alerts. Future work.
How do you know 20+?
Resume.
They already heard Argus Scale
Ylogx ALB / ECS / CloudFront / Docker with RLS already on — more desired count after pool and tier-safe cache. I did not shard. Backup, not a new Ylogx first story.
Keep straight
73% mAP is not shippable. Scale after 89% mAP and 24 FPS.
mAP is held-out detector quality. −50% / 2× are site outcomes after alerts existed.
IQVIA 200+ sites is Invent coverage / Think Big backup, not Success and Scale first story.
Ylogx ECS is Success and Scale backup. Ownership / Frugality / Dive Deep are the Ylogx first stories.
Argus detector is 24 FPS. Horizon camera pipeline is 60 FPS.
Tighten
Gate, then 20+ cameras, then stop. Honest next step is workers or batching, still behind 89% and 24 FPS.
What would you do differently
Follow-up, not a standalone parent stem. Stay on the story just told. If they insist without a parent, Argus per-camera error budget — future work, not a shipped claim.
What they might ask
What would you do differently today?
What would you do differently? (after a deep dive)
IE-asked — follow-up, not a standalone parent stem. Resume-derived IE.in 2024-grad BR list: “What would you do differently today?” inside the deep-dive prompt — IE.in 2024-grad. igreaper BR follow-up list: igreaper (OA-as-R1 on that loop). Rushikesh probes on LC 8362604.
What you say — default if they insist without a parent
Situation. After Argus passed 89% mAP and 24 FPS, I still scaled more than twenty cameras on one global gate.
Task. Name the next control that would have made scale safer, without pretending I already shipped it.
Action. I would add a per-camera error budget so a noisy camera cannot train the floor to ignore alerts. Keep the 89% and 24 FPS bars. Store events, not video blobs. Drop stale frames. That budget is future work.
Result. Same detector, tighter scale. Quality still comes first.
They might ask next
If they just told
You answer
Ylogx Redis + RLS
Denied-tier tests and cache invalidation next to RLS from day one.
IQVIA Hybrid RAG / LangGraph
Freeze gold questions and a citation-required writer earlier.
Horizon ERC
Name the ship-blocker in week one and time-box extra sensor versus fusion.
Argus 73% / 20+ cameras
Per-camera error budget. Future work. Keep 89% and 24 FPS.
GiftedBooks RAG
A small grounded-versus-wrong-citation check. LangSmith is IQVIA, not a GiftedBooks claim.
How do you know the original metric?
Repeat the mechanism plus the resume number. Stay on that product.
Keep straight
Differently is future work on the same product. It is not a new intern and not an invented miss.
IE.in 2024-grad “differently” sits on the deep-dive stem — if that deep dive was Ylogx Redis+RLS, stay there.
Practice plots from the old practice chapter (missed SLA, manager fight, hero weekend) are not yours.
Tighten
Stay on the story just told. One next control, labeled future work, then stop.