10. Behavioral — motivation, process, questions for them
This is a rehearsal notebook for Adarsh Vishwakarma. Read it as yourself. R1 was 13 Aug 2026. R2 was 18 Aug 2026. R3 is done. This folder is leftover Round 4 only. Live Code is Java. Job 10454435 still has no public write-up that names a live-round question. Other people’s stems show shape. They are not a list for this Job ID. Unnamed stays unnamed.
Leadership Principle questions are STAR mapped to Amazon’s 16 official Leadership Principles. Those live on 9. LP list and the answer cards. This page is why Amazon, the intro, the resume walk, current role, GenAI habits, live issues, and questions for them. If the stem is actually a time you received feedback, missed a deadline, took extra work, disagreed, dove deep, exceeded expectations, or took a calculated risk — do not rewrite the STAR here. Open the LP card.
Click a stem in the list. Land on that card. What they might ask is the wording. What you say is one spoken block. They might ask next unpacks the same story. Keep straight is which number lives where. Tighten is how to stop. Spoken STARs for named Leadership Principles live on chapter 9 and in chapter 3. The 60-minute spine, 90-second intro, and close questions also live in chapter 2 — this page must not contradict that hour.
What they might ask
Tell me about yourself. Why Amazon. Walk me through your resume.
A process question: how you work, how you use models, what went live.
“Do you have questions for me?” in the last few minutes.
IE-asked as leftover-hour shape (GFG 2025 current role; LC 7623949 intro then Why Amazon). Resume-derived for the words you speak. Other loop, not a forecast for Job 10454435.
What you say
I answer the stem they asked. If it is a time I received feedback, missed a date, took extra work, disagreed, dove deep, exceeded a bar, or took a calculated risk, I open the LP card and speak that STAR once. I do not restart the same project under a new poster name in the same hour.
Handwritten STAR room rules
Pick a failure that had a real cost, then the action and what you shipped after. Argus 73% mAP, the IQVIA DOM demo, and the Ylogx client-side aggregation crash exist. You do not say you had no significant failures.
Do not reuse the same spoken story for every prompt in the same hour.
Answer the question they asked. Do not pause to name the poster Leadership Principle.
Do not overthink which LP the hour “is.” Do the stem in front of you.
Prepare questions for them at the end. Failing to prepare that close is a handwritten miss. Cards: chapter 2 and questions for them.
They might ask next
If they ask
You answer
They never name an LP
Default pair: Dive Deep (Ylogx Redis and RLS together), then Deliver (Horizon 17th / 80+). Third: GiftedBooks hours → 3–10 min. Full STARs: q-dd-redis-rls, q-dr-horizon, q-co-beyond.
They paste code after a behavioral opener
Leave the story. Write Java. Chapter 6. Fluency-labeled Distance K: Distance K.
They already used that project this hour
Use the named backup on the LP spread — a second story, not a fourth trophy on the same intern.
Keep straight
This page is not sixteen Leadership Principles recited as a speech.
Job 10454435 still has no public named live-round question.
Internships are a timeline: Horizon February–June 2024, Ylogx November 2024–October 2025, IQVIA April 2026–present. Longest production intern is a date: Ylogx, those eleven months.
Tighten
One spoken answer, one project. Stop after one resume family. If the stem belongs on chapter 9, click through instead of inventing a twin STAR here.
What is an LP question vs this page
Same leftover hour. Two kinds of prompt. Mix them in your head and you either dump a STAR into Why Amazon, or you give a process speech when they asked for a time.
Leadership Principle
A “tell me about a time…” or a named principle from the official 16. Situation, task, action, one resume family. Click 9. LP list → answers A (Customer Obsession through Frugality) or answers B (Earn Trust through Success and Scale).
This page
Motivation, resume walk, current role, stack-why, I versus we, GenAI habits, who you ask when blocked, what went live, hypothetical “how would you,” questions for them. Still later-hour talk. Not a second copy of the LP STAR.
What they might ask
Critical feedback, missed deadline, extra task, disagreement with a manager — those sound behavioral. They are LP families.
Split is rehearsal hygiene. Other loops named both kinds in one hour. Not a forecast for Job 10454435.
What you say
If they ask for a time, I speak the locked STAR on chapter 9. If they ask why Amazon, how I work, or what I do now, I stay on this page.
Judgment under a clock, not childhood worship and not a fake team name. This is a method answer, not a trophy STAR. Spoken paragraph matches chapter 2.
IE-asked LC 7623949 leftover: Introduction then Why Amazon then named LPs. IGotAnOffer lists Why Amazon as a common opener — Class C as shape, not a guarantee. Other loop, not a forecast for Job 10454435.
What you say
Situation. I am moving from internships into SDE I, where you own a slice and the clock is real.
Task. Say why this company in intern-honest terms: customer-obsessed engineering, with Leadership Principles as how I already ship — not a poster recitation.
Action. At IQVIA I learned to stop a fluent answer that has no overlapping source span. At Ylogx I learned a SQL chatbot is not done until the database role protects the rows. I want that ownership-and-depth bar at SDE I scope on this AUTA APJ loop.
Result. Three sentences on how I already ship. Then I stop. I do not name a team Amazon has not assigned me.
If you need a personal org/team reason: [YOU FILL: team/org reason — do not invent in the room]. This page does not mint one. No team name is on the resume.
They might ask next
If they ask
You answer
Deep Dive / Learn / Earn Trust after Why Amazon
Ylogx Redis + RLS; IQVIA LangGraph; GiftedBooks grounded RAG + 99.5%. Cards on answers B and answers A.
Why not stay at IQVIA
Intern scope versus SDE I ownership. No insult to IQVIA.
Why this Job ID / Karnataka / AUTA
This is the SDE I loop I am in: AUTA APJ, ADCI Karnataka, Zoom + Live Code. I do not invent a product team.
73% mAP is not a floor alert. Gate is 89% mAP and 24 FPS, then 20+ cameras. That is Argus — only if they ask for a third example. Why Amazon stays the two intern sentences above.
Keep straight
No team, org, or childhood Amazon story on the resume. None in this answer unless you filled the bracket above before the hour.
IQVIA stays a place you learned the span-stop. It is not a company you talk down.
An AWS bill is not on the resume.
Why SDE I versus research scientist / SDM is the next card if they split the stem.
Tighten
Three sentences, then wait. If they name an LP after this, use the pivot table — do not retell Why Amazon as a STAR.
Tell me about yourself / 90-second intro
Current work, three internships as a timeline, then a door they can pick. Current role is IQVIA, April 2026–present. Compact copy of chapter 2 intro. If they cut you at forty-five seconds, name + CUSAT + IQVIA already stand.
IE-asked GFG 2025 leftover opened on current role; LC 7623949 leftover opened on introduction. Resume-derived words. Other loop, not a forecast for Job 10454435.
What you say
I am Adarsh Vishwakarma. I finish a B.Tech in CSE at CUSAT in May 2026 (CGPA 8.42). I am currently a Software Developer Intern at IQVIA in Kochi, April 2026 to present, on two FastAPI systems: LangGraph Deep Research across 200+ websites, and Hybrid RAG on 200+ page BRDs with Azure AI Search, GraphDB, and LangSmith. Before that, at Ylogx from November 2024 to October 2025 I shipped production BI: reports, SQL RAG, three-tier RLS, Redis, and AWS. Earlier, from February to June 2024, I was core software on Team Horizon’s ROS2 rover, 17th of 80+ at ERC 2024. I can go deep on IQVIA retrieval, Ylogx isolation, or Horizon’s live costmap. Which do you want?
They might ask next
If they ask
You answer
The second IQVIA system in more detail
Hybrid RAG on Azure AI Search — hybrid and semantic retrieval plus a GraphDB — for two-hundred-page BRDs, with LangSmith tracing and generated tests. Still IQVIA. Still not Ylogx. Mechanisms: chapter 5.
They cut you at forty-five seconds
The first three sentences already stand. Horizon depth, Argus, GiftedBooks, StratifyLabs, and eight hackathons wait.
Ylogx, November 2024 to October 2025. That is a date.
Keep straight
One intro, three internships as a timeline. StratifyLabs, GiftedBooks, and Argus are projects; they wait until they pick a project.
Ylogx 40 / 99.9 / +65 / −35 / 30 KPI / sub-210 stay off the intro. Those families belong on Ylogx stories.
17th/80+ is in the intro as Horizon’s one line. A full rover STAR waits until they pick that door.
No Amazon team name on the resume. None in the intro.
Tighten
End on “Which do you want?” and wait. Current-role-only wording is the next card if they skip the timeline.
Walk me through your resume
Horizon → Ylogx → IQVIA, then projects only if they pick. This is chronology. It is not a ranking of internships. Same timeline as the home page and chapter 2.
Resume-derived dates and jobs. Other loop, not a forecast for Job 10454435.
What you say
Situation. Three internships at three times. Each job is a different product.
Task. Give dates and what the job was, then stop unless they pick a door.
Action. Horizon was February to June 2024: rover software under an ERC date that does not move. Ylogx was November 2024 to October 2025: a year of production BI — reports, SQL RAG, three-tier isolation in Postgres, Redis, AWS. IQVIA is April 2026 to present, Kochi: my current role, retrieval and evals on Deep Research and Hybrid RAG.
Result. Three jobs in order. Projects (Argus, GiftedBooks, StratifyLabs) wait until they ask. Education is CUSAT CSE, May 2026, CGPA 8.42, IEDC Tech Team member, eight hackathons — only if they still want the rest of the page.
They might ask next
If they ask
You answer
Pick one intern and go deep
That product’s card in chapter 5. One system. One resume family.
They point at Argus / GiftedBooks / StratifyLabs
Those are projects, not a fourth internship. Speak that product only.
Which intern ran longest in production
Ylogx, November 2024 to October 2025.
They open GitHub
Resume wins. GiftedBooks README currently describes AegisAI. StratifyLabs GitHub is a Next.js scaffold. Chapter 8.
Keep straight
Dates do not overlap. Horizon ends June 2024. Ylogx starts November 2024. IQVIA starts April 2026 after Ylogx ends October 2025.
Horizon is a competition rover, not a SaaS internship.
Ylogx chatbot is LangChain SQL RAG. LangGraph and LangSmith are IQVIA.
99.9% and sub-210 ms are Ylogx. IQVIA has no uptime or latency percent on the resume.
Tighten
Stop after the three jobs unless they pick one door. Why-three is the same timeline on that card — do not rank the internships if they ask why there are three.
Current working role and responsibility
Different stem from tell-me-about-yourself. Still not an LP dump. Two FastAPI systems. Resume numbers only. Stop. Matches chapter 2 current role.
Current working role and responsibility (GFG 2025 leftover, printed).
What do you do right now?
Walk through current project architecture.
IE-asked GFG 2025 leftover hour. LC 6866029 leftover-shaped architecture stem. Resume-derived. Other loop, not a forecast for Job 10454435.
What you say
Situation. I am a software developer intern at IQVIA in Kochi, April 2026 to present.
Task. Two retrieval systems analysts can trust: research across the web, and questions over long BRDs.
Action. I work on two FastAPI systems. The first is LangGraph Deep Research across more than two hundred websites. The second is Hybrid RAG: Azure AI Search hybrid and semantic retrieval, a GraphDB, LangGraph choosing retrieval, LangSmith for evals and generated test cases on more than two-hundred-page BRDs.
Result. Two 200+ families: websites ranked, page BRDs processed. The resume has no IQVIA latency or accuracy percent. Then I stop.
They might ask next
If they ask
You answer
What did you do before
One sentence Ylogx (November 2024–October 2025: FastAPI, NestJS, Postgres, SQL RAG, three-tier RLS, Redis, ECS) or one sentence Horizon (February–June 2024: ERC 17th of 80+). Then wait.
Whether this intern is Java
Production here is Python FastAPI. I write Java in Live Code.
Offer one failure mode: two high-ranked pages disagree, or hybrid versus vector-only. Architecture retell waits. Invent/Learn/Backbone STARs are chapter 9, not a fourth IQVIA trophy in this opener.
Keep straight
GraphDB stores requirement dependencies. LangGraph is workflow state.
Spring is not on the IQVIA bullets. Production there is Python FastAPI.
−35% bot DB and sub-210 ms are Ylogx. IQVIA has no latency percent on the resume.
Current role is not Ownership / Invent / Learn spoken as three IQVIA trophies in the intro.
Tighten
Stop after two systems. If they want a principle, chapter 9. If they want mechanisms, chapter 5.
Best project and worst / hardest project
Best = one project with a reason about the work. Worst / hardest = a real cost, then what you changed. You do not trash an employer. Two spoken blocks because they are two prompts.
What is your worst project / most challenging project / a failure?
IE-asked most-challenging / failed in other leftover hours (LC 7035289, LC 6866029 as shape). Resume-derived costs. Other loop, not a forecast for Job 10454435.
What you say
Best — Situation. If they let me pick one, I pick Ylogx isolation. Operators shared warehouse facts across three organizational tiers.
Task. A SQL chatbot that returns rows is not done if a JOIN can cross a tenant.
Action. I put three-tier RLS and RBAC in Postgres, executed generated SQL as the caller’s non-owner role, and put the tenant tier in the Redis key.
Result. Bot database latency dropped 35 percent. Uptime on that product is 99.9 percent. The reason I call it best is the gate: legal SQL can still be the wrong tenant. That is a reason about the work, not a ranking of internships.
Worst / hardest — Situation. On Argus the first honest detector sat at 73 percent mAP.
Task. Floor alerts at that quality miss violations and also propose junk boxes. Either one trains people to ignore alarms.
Action. I treated 73 percent as not shippable. I iterated data and configuration on 15,000+ images to 89 percent mAP, verified 24 FPS end-to-end, then scaled to 20+ cameras.
Result. 73→89 percent mAP, 24 FPS, then 20+ cameras. After alerts existed, the site saw −50 percent violations and 2× compliance — those last two are plant outcomes, not mAP. The cost was shipping a detector people would ignore. The change was the gate.
They might ask next
If they ask
You answer
Best, but they already heard Ylogx
Horizon under an ERC date that does not move, or IQVIA retrieval with cite-or-abstain. One reason, one project. Not a contest speech.
Hardest as IQVIA instead of Argus
Stakeholder demo: DOM truncated, agent clicked the last chart zoom. Full STAR: q-dd-dom. No IQVIA percent.
Hardest as a Ylogx crash
Client-side aggregation on a large report slice crashed the UI. Server/preload path after. Notes STAR: q-hs-crash. 40% / 99.9% are the product family, not a claim that this crash produced those percents alone.
They want a time you were wrong
Argus 73% is also Are Right: q-ar-mistake. Do not start a second failure in the same hour.
Was Ylogx or IQVIA a bad employer
No. Worst/hardest is a defect class you owned, not a review of the company.
Keep straight
73% mAP is Argus. It is not a shippable floor alert. Gate is 89% mAP and 24 FPS, then 20+ cameras.
−50% / 2× are site outcomes after alerts. They are not mAP.
99.9% is Ylogx. IQVIA has no uptime percent.
Missed ERC is not on the resume. Hardest Horizon beat is sequencing risk (stale occupancy), not a fake miss.
Tighten
Speak one of the two blocks — the one they asked. If they ask both, best then worst, then stop. If Argus already appeared this hour, worst switches to IQVIA DOM or Ylogx crash.
How did the idea come about? Stack and why? Challenges?
Point at one project. Default is current IQVIA. If they named Ylogx or Horizon, stay on that product. Compact why. Mechanisms live in chapter 5 — do not retell the whole data flow unless they ask.
IE-asked project-genesis / stack-why as leftover shape in other loops. Resume-derived mechanisms. Other loop, not a forecast for Job 10454435.
What you say
Situation. At IQVIA, April 2026 to present, analysts cannot treat the first web hit as truth and cannot live inside 200+ page BRDs by stuffing a PDF into a context window.
Task. Two FastAPI systems with the same gates: stateful orchestration, isolated tools, cite-or-abstain, a trace that fails if the supporting span is missing.
Action. Why LangGraph: a linear scrape-then-write chain loses Bing hits when Playwright times out, and there is no explicit edge for citation failure to return to retrieval. Typed state, parallel tools, append-only reducers, checkpoints, and a bounded retrieve-grade cycle isolate those failures. Why Azure hybrid plus GraphDB: vector-only misses rare clause IDs such as REQ-1044; keyword-only misses paraphrased policy; “what breaks if this requirement changes” is a dependency edge, not a cosine neighbor. Hybrid keeps exact tokens beside paraphrases. GraphDB walks stored depends-on edges. Neighbor text still comes from Azure. LangSmith gold items expect citation IDs. No supporting span, no sentence.
Result. Deep Research ranked across 200+ websites. Hybrid RAG on 200+ page BRDs. No IQVIA latency percent. Non-technical challenge: a professional paragraph can still cite the wrong annex — that is why the writer may abstain. Then I stop and offer one door: Deep Research, Hybrid RAG, or evals.
They might ask next
If they ask
You answer
They named Ylogx
Why RLS in Postgres: an app-only WHERE dies on the first missed JOIN or generated SQL that skips the helper. Why Redis: every NL turn repeating schema and hot analytics is felt as “the model is thinking”; a bigger RDS repeats the same work; keys carry tier and role. Dashboards have no LLM; sub-210 ms is that GET. Full path: chapter 5 Ylogx.
They named Horizon
Why GStreamer 60 FPS: operators need the newest frame. Why fuse sensors we had: extra hardware near ERC adds drivers and delay while occupancy stays unsolved. chapter 5 Horizon.
Walk the boxes
Stay on the product they named. Numbered flow is chapter 5, not a second intern.
LangGraph and LangSmith are IQVIA. Ylogx chatbot is LangChain SQL RAG.
GraphDB is requirement-dependency storage. LangGraph is workflow state.
Redis −35% and RLS three tiers are Ylogx. They do not belong in the IQVIA spoken block.
FAISS is a skills line. IQVIA intern retrieval is Azure AI Search plus GraphDB.
Tighten
One product. Two “why this mechanism” sentences. Then wait. Chapter 5 if they want the numbered path.
Your contribution versus the team
I for modules you shipped. We for Horizon 17th / 80+, site −50% / 2× after Argus alerts, and named team hackathon prizes. Intern title stays intern. Lookup also in chapter 8.
IE-asked “what did YOU do” as later-slot shape. Resume-derived ownership. Other loop, not a forecast for Job 10454435.
What you say
Situation. They want the module, not the company name on the slide.
Task. Name what I implemented. Name what is a team score.
Action. I was an intern. I put three-tier RLS in Postgres on the shared chatbot and report path. I measured bot database latency and put the tenant tier in the Redis key.
Result. Bot database latency dropped 35 percent. That sentence is I. Then I wait. Horizon 17th is we — that row is in the table if they switch products.
IQVIA two FastAPI systems, LangGraph nodes, LangSmith fail-if-no-span
I as intern. Not Head of Platform.
Argus train / mAP–FPS gate / NMS / event log
I. Site −50% / 2× involve the plant after alerts existed.
Horizon GStreamer, ZED mapping, costmap software
I owned the path. Chassis / electrical: not claimed as mine.
ERC 17th / 80+
We
CodeRecet 1st, MLH Best Project, Magnathon 2.0 runner-up
We for team prizes. I paired and shipped modules.
IEDC Tech Team
I shared setup as a member. No mentee count. No hire/fire.
Keep straight
Architected / Built / Designed on the resume still unpacks into modules an intern implemented.
Plant headcount is not on the PDF. −50% / 2× stay we-shaped site outcomes.
I do not I-wash 17th.
Tighten
One module per sentence. If they ask what the team did: ERC 17th and named prizes are we. Isolation, gates, and camera software stay I. Then stop.
What would you do differently
Same-path missing test — not a new intern, not a fake missed ERC. Default to the project they already opened. If they ask cold, Argus 73% gate earlier.
IE-asked 3–4 levels of why / differently (igreaper leftover shape). Resume-derived and Notes for DOM / crash. Other loop, not a forecast for Job 10454435.
What you say
Situation. On Argus I measured 73 percent mAP before I treated it as a product.
Task. The miss is sequencing: the gate should have been explicit before anyone talked about cameras.
Action. What I would sequence earlier is the held-out mAP plus 24 FPS bar as a written stop, before a demo that looks live. Same habit I later used: 89 percent and 24 FPS, then 20+ cameras.
Result. The resume already has that gate. Differently means I would have named the stop in week one of the detector, not after the first honest eval surprised me. Then I stop.
They might ask next
If they ask
You answer
Same path is IQVIA
Raise DOM max earlier (50 → 80 elements, reasoning med) before the stakeholder demo. Freeze gold questions and citation-required writer earlier. q-dd-dom.
Same path is Ylogx
Denied-tier tests beside the RLS migration, not a bigger cache first. Client aggregation: keep heavy aggregation off the browser when the slice is large — q-hs-crash.
Same path is GiftedBooks
Fix the README so it matches the resume product. GitHub currently describes AegisAI. Resume wins. q-et-github.
Same path is Horizon
Name the ship-blocker in week one: a healthy 60 FPS feed with a stale costmap is not a delivered rover. ERC was not missed. I do not invent a miss to sound humble.
Keep straight
Missed ERC and missed 99.9% are not on the resume. Differently is a missing test on a path you shipped.
Stay on the project they already heard unless they change products.
Tighten
One missing test. Then wait. This is not a second Are Right STAR unless they ask for a time you were wrong — then q-ar-mistake.
Questions for them — last minutes
Handwritten miss: failing to prepare questions at the end. Ask about the work. Do not pitch a product. Compensation belongs in a later recruiter conversation. Same close as chapter 2, with a longer menu so you can pick two.
After two LP sub-questions, the same close (Reddit 1qj304z).
IE-asked end-of-hour shape. Reddit 1qj304z: candidate pitched a Prime Video rating idea; the interviewer would not discuss it; hour ended ~25 minutes. That reverse-Q is a warning, not a template. Other loop, not a forecast for Job 10454435.
What you say
Situation. They offer the last few minutes.
Task. Ask about the team’s week, not the loop’s verdict.
Action. I pick one or two from this list, listen, stop:
What does a typical week look like on this team for a new SDE I?
How does a new SDE I pick up an existing service — docs, pairing, first change in production?
What does on-call look like here in the first months?
How does this team review AI-generated code — what still has to be a human check?
What problem space does this org actually build — the work a new SDE I would see, not a confidential roadmap?
What separates someone doing well at six months from someone who is not?
What have you seen new grads struggle with on Live Code versus in production?
What has kept you at Amazon?
Result. I learn how the team works. I do not extract a hiring signal. I do not pitch a product.
They might ask next
If they ask
You answer
They answer with a team problem
One follow-up on on-call or on how SDE I code review works. Then stop.
They say they cannot share
Thank them and stop. Confidential roadmap was never the ask.
Time remains and they open an editor
Switch to Java. Shiwangi’s leftover DSA existed because time remained.
They end early
Questions from this list if offered, thank you, stop. Early end is not a pass/fail script. Chapter 2 early.
Keep straight
Compensation, levels, and competing offers wait for the recruiter conversation.
“What does Amazon do” is not a question for this hour.
Prime Video, Alexa, or a startup pitch is not the close.
Whether this was the Bar Raiser is not a question you ask.
Tighten
One or two questions, listen, stop. If they open an editor, chapter 6.
How you use GenAI, limitations, hallucinations, how you validate
Verify, do not worship autocomplete. Resume-true: IQVIA LangSmith and cite-or-abstain; Ylogx SQL RAG constrained, never superuser. If they paste DSA in a Fluency slot, write Java first. Full walls in chapter 7. Distance K: chapter 6 and chapter 7 Fluency tree.
How do you use GenAI in projects? Where should we use them and where not?
Limitations / hallucinations / how do you validate?
A slot labeled Gen AI Fluency that is nodes at distance K, no parent map.
IE-asked LC 7850431, 7724048, 8014509; Fluency-as-tree LC 7623949. Resume-derived walls. Other loop, not a forecast for Job 10454435.
What you say
Situation. A fluent paragraph can still be wrong. A Fluency label can still be a tree.
Task. Use a model where a human still checks a citation or a SQL shape. Do not use it where a wrong output is a tenant leak.
Action. At IQVIA I draft with a model behind hybrid retrieval. I verify with LangSmith traces and a required overlapping span. No supporting span in retained chunks means no sentence — cite-or-abstain. Empty scrape is not a fact. At Ylogx, LangChain SQL RAG may propose constrained NL → SQL only. Postgres executes as the caller’s non-owner role. RLS still filters. Catalog and writes are rejected. The model does not author row policy or sit on the dashboard GET.
Result. IQVIA: research ranked across 200+ websites; Hybrid RAG on 200+ page BRDs. Ylogx SQL RAG is the +65% analysis path. Dashboards are sub-210 ms with no LLM on that GET. If the editor is open, this speech waits — I write Java first.
They might ask next
If they ask
You answer
They pasted Distance K / a tree
Write the Java. Connect children, BFS from target to distance K, or the no-parent-map walk in chapter 6. One sentence after code: I would not put an LLM on this walk.
How you verify generated code
Treat it as untrusted. Compile, dry-run their example, complexity, forbidden structures (parent map). The interview is the Java I type. Chapter 7 verify-code.
Hallucinations on IQVIA
Two high-ranked pages disagree: ranker must not average them into fake agreement. Writer abstains or shows the conflict. LangSmith gold expects citation IDs.
Could the model write RLS?
HTTP 200 with a legal JOIN is still a leak. Engine is Postgres RLS as the caller role. Full walls: chapter 7 must-not.
GiftedBooks
Answers stay grounded in that student’s PDF. sub-300 ms is that API. README currently describes AegisAI — resume wins if they open GitHub.
Keep straight
LangSmith is IQVIA. Ylogx chatbot is LangChain SQL RAG, not LangGraph.
99.9% and sub-210 ms are Ylogx. IQVIA has no uptime or latency percent.
GraphDB ≠ LangGraph. Sharing the word “graph” is how this answer dies.
Intern Python sidecar is not Live Code. Live Code is Java.
Tighten
One intern path if they ask how you use models. Tree first if they pasted a tree. Empty editor means no LangGraph lecture.
How you approach a complex problem
Process, not a second Dive Deep STAR. Clarify → options → smallest reversible test → measure. If they want a time, open q-dd-pocs or q-dd-dom.
Class C process prompt. If they add “tell me a time you researched POCs,” that is Dive Deep on chapter 9. Other loop, not a forecast for Job 10454435.
What you say
Situation. The prompt is how I work, not yet a named principle.
Task. Get to a reversible test before a rewrite.
Action. I clarify what “done” means and what must not break. I name two or three options, including the obvious one that fails. I pick the smallest reversible test I can measure. I read the result, then I commit or I change the option.
Result. The next sentence is a measurement, not a slogan. If they want a time: IQVIA hybrid versus vector-only as researched options — q-dd-pocs — or the DOM demo when the agent clicked the wrong zoom — q-dd-dom.
They might ask next
If they ask
You answer
Give an example anyway
One pointer. Default q-dd-pocs if Invent/Dive Deep unused. If Ylogx Dive Deep already ran, q-dd-dom. Do not paste the STAR here.
How do you know the test worked
IQVIA: gold citation IDs / span overlap. Ylogx: deny fixture, zero foreign rows. Argus: held-out mAP, not a caption.
Class C process. Notes for the IQVIA demo logs habit. Other loop, not a forecast for Job 10454435.
What you say
Situation. I am blocked on a path I own — intern title, still the module.
Task. Unblock with evidence, not with a skip-level meeting.
Action. First I read the failing logs, traces, or deny fixture myself. Then I ask the person who last touched that module, or I read the runbook / docs, with a specific question and what I already tried. I do not open with “it is broken.”
Result. The next action is a smaller test. On the IQVIA stakeholder demo I used logs, audit logs, reports, and screenshots the run took — then raised DOM max. Full time: q-dd-dom.
They might ask next
If they ask
You answer
What if the teammate is wrong
Bring the fixture. Hybrid versus vector is Backbone: q-bb-disagree. Hypothetical method: that card.
Ylogx version
Cross-tier deny fixture and the JOIN shape, then the person who owns schema — not a skip-level.
Did you escalate to leadership
Intern path. Logs and the owning teammate. I do not invent skip-level theatre.
Keep straight
A manager fight is not on the resume.
IEDC is Tech Team member. No mentee count. Help looks like pairing, not skip-level.
Tighten
Logs, then a specific question to a teammate. Stop. If they want the demo STAR, chapter 9 — do not retell it here.
How you divide an ambiguous problem
Default spoken path is IQVIA: planner / ranker / writer as typed stages. If they are already on Horizon, date versus software fusion. One project in the spoken block.
Resume-derived IQVIA nodes and Horizon date. Other loop, not a forecast for Job 10454435.
What you say
Situation. At IQVIA a research query does not arrive as a finished retrieval plan. Two credible pages can disagree. An empty scrape can still look like a fluent draft.
Task. Split the work so each stage can fail without poisoning the next.
Action. Planner writes a search plan into LangGraph typed state. Tool nodes fan out. Ranker scores source quality, recency, and agreement — syndicated mirrors are one vote. Writer synthesizes only from ranked evidence or abstains. A citation grader may hop back with an iteration cap. LangSmith traces node order so an empty scrape reaching the writer is visible.
Result. Research ranked across 200+ websites. Ambiguity became stages with a stop, not one giant prompt. No IQVIA percent. Then I stop.
They might ask next
If they ask
You answer
Horizon instead
The date does not move. I sequenced GStreamer 60 FPS and the live costmap first. Extra hardware was the cut. Fusion on sensors we had. 17th / 80+ is we. Bias/Deliver STARs: q-bfa-date, q-dr-horizon.
Lexical versus semantic versus bounded GraphDB hop. Writer sees a small evidence set, not 200 pages. Chapter 5.
Keep straight
Planner / ranker / writer is IQVIA Deep Research. SQL RAG planner language is Ylogx — different intern, different library (LangChain).
200+ websites is Deep Research. 200+ page BRDs is Hybrid RAG.
Tighten
Name the stages. Stop after one 200+ family. If IQVIA already appeared, use Horizon date versus fusion — still one project.
Collaboration / working with a client / demo
Client-demo beat is the IQVIA stakeholder demo (DOM truncation). Compact here. Full STAR: q-dd-dom. Hackathons: pitching and feature-versus-deadline notes — not invented hours-to-prize. Showcase pair also on client demo.
Tell me about working with a client / stakeholders.
Tell me about a demo that went badly, then worked.
How do you collaborate on a student team / hackathon?
Notes IQVIA DOM demo. Resume-derived hackathon placements. Other loop, not a forecast for Job 10454435.
What you say
Situation. At IQVIA the team had to demo an AI application stakeholders were using to test the platform.
Task. The run failed in front of people who were judging the product, not the prompt poetry.
Action. The agent clicked the last chart’s zoom instead of the intended chart’s zoom. I was confused until logs, audit logs, reports, and screenshots showed DOM truncation — the agent could not see the right control, assumed the last matching element, and clicked wrong. Prompt/reasoning was also too thin (low did not work) and the UI clipped. I raised DOM max characters in config, max DOM elements 50 → 80, reasoning to med.
Result. The path worked. We demoed to stakeholders. The resume has no IQVIA percent — I stop there. Full spoken STAR is q-dd-dom.
They might ask next
If they ask
You answer
Hackathon collaboration
IEDC Tech Team member. I paired until someone else could launch. Pitching and a weekend clock: feature versus deadline — ship a demo someone else can run, not a hero laptop. Named placements are we: CodeRecet 1st, MLH Best Project, Magnathon 2.0 runner-up. No hours-to-prize story on the resume. Hackathons card.
Ylogx “client”
Operators and non-SQL users on shared warehouse facts. Report stability: client demo and q-hs-crash.
They want Dive Deep logs
This is that demo. Use q-dd-dom. Do not start Redis in the same hour unless they change products.
Keep straight
DOM 50 → 80 and reasoning med are handwritten notes. No IQVIA latency or accuracy percent.
Stakeholder demo is IQVIA. ERC field test is Horizon. They are not the same client story.
Tighten
One demo path. If they already heard IQVIA, switch to Ylogx report stability on the client-demo card — not a fourth IQVIA primary.
What issues when the project went live
Ylogx, November 2024–October 2025. Speak failure classes, not a fabricated ticket. Dev → QA → prod is blast radius. Dashboards are sub-210 ms without an LLM on that GET. Companion: older chapter 15 and r4 chapter 5.
What issues came up when you went live / in production?
How do you promote from dev to QA to prod?
How does RLS / RBAC actually work?
Resume-derived isolation, Redis, dashboards, 99.9%. Other loop, not a forecast for Job 10454435. No named incident on the resume.
What you say
Situation. At Ylogx, operators shared warehouse facts across three organizational tiers. Generated SQL and cached answers could look green while being wrong for the tenant.
Task. Promote only after the class of leak is denied at the smallest blast radius.
Action. Dev: non-owner SQL, cache miss plus Redis fallback, KPI GET without the model. QA: allow and deny fixtures across three tier shapes, denied JOINs, golden NL → SQL shape, wrong-role cache keys, Redis-down fallback, all 30 dashboards without generation. Prod: the same gates, larger blast radius. A new image tag is not a safety proof.
Result. Resume families on this intern: 40% faster reports, 99.9% uptime, SQL RAG +65%, Redis −35% bot DB, 30 dashboards +60% ops, sub-210 ms. I pick one class if they want an example, then I stop.
Three live classes (pick one)
Class
What it looks like
What you changed
JOIN leak
Legal SQL, HTTP 200, another organization’s rows through a mapping table
Caller non-owner role; RLS on joined protected tables; deny fixture asserts zero foreign rows
Wrong-tier cache
Identical English, “top SKU,” shared Redis key across tiers
Key = tier, role, query fingerprint, schema version. Redis down: same caller role, never owner
LLM on dashboard GET
Thirty tiles become thirty generations; KPIs become inventable; sub-210 ms dies
Deterministic NestJS REST + Postgres RLS + Recharts. Ad-hoc English stays on SQL RAG
RLS and RBAC in one table
Piece
What it decides
NestJS RBAC
Whether the caller may hit the route (view dashboard, build report, call chatbot)
Postgres RLS
Which rows exist for that role — ordinary reads and generated SQL
Caller DB role
Mapped identity, never table owner / superuser
Redis key
Tier and role in the key so a cached answer cannot leak
Dashboard GET
No LLM. sub-210 ms lives here. −35% lives on the bot DB path
They might ask next
If they ask
You answer
Was this a named outage
No. These are classes the architecture prevents. No ticket, customer, or duration on the resume.
Dive Deep on this intern
q-dd-redis-rls — Redis and RLS together. Ownership q-own-extra if they ask extra work. Do not spend a fourth Ylogx LP trophy.
Client aggregation crash
q-hs-crash. Notes. 40% / 99.9% are the product family, not this defect alone.
10×
More RLS and deny fixtures, role-correct pools, then more ECS desired tasks. I did not shard Postgres or Redis.
Keep straight
−35% is bot database latency via Redis. sub-210 ms is dashboard serving with no LLM on that GET. Two families.
99.9% uptime is Ylogx. GiftedBooks is 99.5%. IQVIA has no uptime percent.
Literal tier names and private table names are not on the resume. Speak org-, team-, and user-shaped levels.
Kubernetes, Kafka, and EKS are not on this intern’s bullets. Shipped path: FastAPI, NestJS, Postgres, Redis, ECS, ALB, CloudFront, Docker.
Tighten
One class, then the RLS/RBAC split if they ask. Stop. If Ylogx already filled Ownership, Frugality, and Dive Deep this hour, do not start a fourth Ylogx LP — stay on this process card or switch products.
Eight hackathons and IEDC Tech Team
IEDC Tech Team member. Eight events. Three named placements. Pitching, presentations, feature versus deadline. No invented hours-to-prize. No mentee count. Depth: chapter 5 IEDC. Hire STAR: q-hd-helped.
Resume-derived eight events, CodeRecet 1st, MLH Best Project, Magnathon 2.0 runner-up, CUSAT 8.42. Other loop, not a forecast for Job 10454435.
What you say
Situation. I studied CSE at CUSAT, October 2022 to May 2026, CGPA 8.42. I served on the Tech Team at IEDC CUSAT — a team member. Club and hackathon stacks were not shared. Weekend clocks do not wait for a perfect architecture.
Task. Ship a demo someone else can run. Transfer the risky module. Pitch what actually works.
Action. I shared environment setup — commands and dependencies, not a slide. I paired until another laptop reproduced the launch. Feature versus deadline: cut scope that only worked on my machine. Pitching: what the judge can run in the slot, not a tour of five unnamed events.
Result. Eight national and regional hackathons. We placed first at CodeRecet, MLH Best Project, Magnathon 2.0 runner-up. No mentee count. No hours-to-prize. Then I stop.
They might ask next
If they ask
You answer
Hire and Develop
q-hd-helped. Member, not manager. Backup is Horizon pairing until someone else can launch GStreamer / costmap.
Bias for Action and Horizon already ran
Eight named hackathons are the Bias backup. Horizon remains Bias primary if unused.
The other five event names
Not on the resume as names. I do not invent architectures for them.
I versus we on prizes
Prizes are we. I paired and shipped modules.
Keep straight
IEDC Tech Team member. No mentee count. CGPA 8.42 is education, not a coaching metric.
Hackathon placements are event outcomes, not Argus mAP, not Ylogx 99.9%, not a second ERC medal.
Hire primary is IEDC. Earth’s Best primary is Horizon knowledge sharing.
Tighten
Member, pairing, three named events. Stop. If they want leadership as an LP, open Hire — do not dump intern metrics as a prize system.
How we ranked 17th at ERC
We placed 17th of 80+. Your slice is software: GStreamer 60 FPS, ZED 2 at 2M+ points/s, costmap −55%. Core software member, not team lead, not chassis. Full rover: chapter 5. Deliver STAR: q-dr-horizon.
Resume-derived 17th / 80+, 60 FPS, +40% obstacle, −55% collision, ZED 2M+. Other loop, not a forecast for Job 10454435.
What you say
Situation. Team Horizon, Kochi, February to June 2024. European Rover Challenge 2024. The date does not move. A rover that only sees pixels still collides. A live video feed is not a costmap.
Task. Current operator camera, mapping that keeps local density, fused occupancy the planner can trust.
Action. I owned software around the camera pipeline, ZED mapping, and costmap. GStreamer at 60 FPS — drop stale packets rather than stall later bytes. ZED 2 at 2M+ points per second incoming; planning keeps local density. Fusion writes occupied, free, and uncertain cells. Under the date we fused sensors we already had.
Result. Resume: 60 FPS, +40% obstacle detection, −55% collision risk, ZED 2M+. We placed 17th among 80+ teams. Placement is we. I was not team lead. I did not design the chassis.
60 is the rover camera pipeline. 24 is the PPE detector budget. Different products.
Did you miss ERC
No. That miss is not on the resume. Sequencing risk is a stale costmap behind a healthy feed.
Keep straight
+40% obstacle and −55% collision are rover outcomes. Argus mAP and Argus −50% violations are a different product.
Horizon is a competition rover, not a SaaS internship.
No YOLO-on-rover resume claim. No LLM on occupancy.
ERC listed twice on the resume is the same 2024 result.
Tighten
We 17th, then your three software numbers. Stop. If Horizon already filled Bias and Deliver this hour, Earth’s Best is sharing until someone else can launch — or Ylogx 40% + 99.9% as Deliver backup.
Why three internships
Dates and jobs. If they ask which intern ran longest in production, answer with a date: Ylogx, November 2024 to October 2025.
Resume-derived timeline. IE-asked current role in other leftover hours. Other loop, not a forecast for Job 10454435.
What you say
Situation. Three internships at three times. Dates do not overlap.
Task. Say what each job was.
Action. Horizon, February to June 2024: rover software for ERC 2024. Ylogx, November 2024 to October 2025: production BI — reports, isolation, latency, AWS. IQVIA, April 2026 to present: current role, retrieval and evals.
Result. Three jobs on a timeline. If they ask which intern ran longest in production: Ylogx, November 2024 to October 2025. That is a date.
They might ask next
If they ask
You answer
Why leave Horizon / Ylogx
The dates ended. Horizon was a contest window. Ylogx ended October 2025. IQVIA is current. No insult.
Pick one to go deep
The door they name. Chapter 5. Not a ranking speech.
Projects vs internships
Argus, GiftedBooks, StratifyLabs wait until they ask for projects.
Keep straight
Horizon ends June 2024. Ylogx starts November 2024. IQVIA starts April 2026.
This card does not compare internships inside the spoken answer.
Horizon is not a SaaS internship.
Tighten
Three dated jobs. Stop. Resume walk is the same timeline if they want the rest of the PDF.
Why Java in Live Code if internships were Python / TypeScript / ROS2
This loop’s editor language is Java. Internship production is not a JVM rewrite of those bullets. Say the split here. You do not need to restate it in every later answer.
Your resume looks like Python and TypeScript — why Java here?
Have you shipped Spring / Kafka services?
Can you code in Java?
Resume-derived intern stacks vs this Job ID’s Live Code. Recruiter “LP-only” still opened an editor in other loops (LC 7406809, 6806195). Other loop, not a forecast for Job 10454435.
What you say
Situation. Amazon Live Code on this loop is Java. My internships shipped Python FastAPI / LangGraph / YOLOv9, TypeScript NestJS/React, and ROS2.
Task. Close the gap in the editor in front of us.
Action. I restate the problem, name a brute that is too slow, pick a structure, say time and space, and type Java. I do not describe those internships as Spring or Kafka services — those are not on those bullets.
Result. Production intern language stays what shipped. This hour’s language is Java. If they paste a tree, I write the tree.
They might ask next
If they ask
You answer
Something you are not good at
Honest intern gap: no JVM internships; Live Code is still Java. Earn Trust card: q-et-not-good.
What shipped is ECS, Docker, ALB, CloudFront. Kubernetes is a skills-line word, not that intern’s serving path.
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.
Login Tracker is unverified. Rate Limiter independent SDE I live count is 1, not UTA-default leftover LLD.
Tighten
Two sentences on the split, then wait for the editor. Do not apologize through a sixth project.
Why SDE I (not research scientist, not SDM)
You ship isolation, retrieval evals, and date-driven software. Influence without a manager title. Not a research-scientist loop. Not hire/fire or budget — none of those are on the resume. IEDC was Tech Team member.
Resume-derived intern scope. Other loop, not a forecast for Job 10454435.
What you say
Situation. I am finishing internships and applying for SDE I on this AUTA APJ loop.
Task. Own a slice of a service: implement, gate, measure. Not a paper mill. Not people-management.
Action. I already ship that way. Isolation: RLS in Postgres when the chatbot ticket would have stopped at SQL. Retrieval evals: cite-or-abstain and LangSmith so a fluent paragraph without a span does not ship. Real-time software: Horizon under an ERC date. Influence without authority looks like holding a 73% detector until 89% mAP and 24 FPS, and putting row policy in the database instead of hoping every handler remembers WHERE.
Result. SDE I is the title that matches that work. I did not hire or fire anyone. I did not own a budget. IEDC was member. Then I stop.
[YOU FILL: team/org reason — do not invent in the room] or skip. No team on the resume.
Keep straight
This is not sixteen Leadership Principles as a career speech.
73% gate is Argus. RLS-in-DB is Ylogx. Cite-or-abstain is IQVIA. Name the product when you name the gate.
No budget, firing, or skip-level story on the resume.
Tighten
Title, three kinds of shipping, intern-honest authority. Stop. Strengths one-liners live on that card if they split the stem.
How would you handle disagreement (hypothetical)
They have not asked for a time yet. Method: data, smaller test, commit. If they then want a time, q-bb-disagree — IQVIA hybrid versus vector-only. Do not rewrite that STAR here.
How would you handle it if you disagreed with a teammate / a manager?
What would you do if your approach was rejected?
Class C hypothetical. Named leftover hours used “disagreement with manager” as a STAR — that is Backbone on chapter 9. Other loop, not a forecast for Job 10454435.
What you say
Situation. They asked how I would handle disagreement, not yet “tell me about a time.”
Task. Show the method I already use as an intern: evidence, a reversible test, then commit.
Action. I bring a measurement or a failing fixture, not a personality story. I propose a smaller reversible test. If the owner of the path decides after that test, I commit on the eval or the deny fixture — I do not relitigate in chat.
Result. The disagreement becomes a test someone can rerun. If they want a time I already shipped: IQVIA hybrid versus vector-only, then commit on LangSmith evals — full STAR q-bb-disagree. Backup path: Ylogx RLS-in-database versus app-only WHERE.
They might ask next
If they ask
You answer
Tell me about a time
Open q-bb-disagree. Stop the method speech.
What if you are wrong
The test can falsify me. Argus 73% is a time I was wrong about a shippable detector: q-ar-mistake — only if they want that time and Argus is still free.
Manager quote / fight
Not on the resume. I do not invent one.
Keep straight
Backbone primary is IQVIA hybrid versus vector-only. Ylogx RLS-in-DB is the backup.
A manager fight is not on the resume.
Tighten
Method in four sentences. If they want a time, click the Backbone card — do not paste it.
Client demo to showcase
Handwritten blank in notes — fill with IQVIA DOM (stakeholders) and Ylogx report stability. Two spoken blocks. Full STARs: q-dd-dom and q-hs-crash.
Notes both stories. Other loop, not a forecast for Job 10454435.
What you say
IQVIA demo — Situation. Stakeholders were using the AI application to test the platform. The agent clicked the last chart’s zoom instead of the intended chart’s zoom. It also could not compare two charts’ data on the fly.
Task. Debug from logs, audit logs, reports, and screenshots, then make the path demoable.
Action. DOM was truncated, so the agent assumed the last matching element. I raised DOM max characters, max DOM elements 50 → 80, reasoning from low to med.
Result. Path worked; demoed to stakeholders. No IQVIA percent. Full STAR: q-dd-dom.
Ylogx report demo — Situation. Dynamic reports with aggregation and filters across chart types. Aggregation on the client. The UI crashed when a large volume hit the browser.
Task. Keep heavy aggregation off the client when the user asks for a large slice, so the demo stays up.
Action. Server/preload path for the AI-requested report. Explicit user click to load full data. Cleanup so the agent and UI stay stable. Multi-chart generation via the LLM path, not report-as-image.
Result. UI rendered; agent stayed stable. Resume family on this product: 40% faster reports and 99.9% uptime — those percents are not claimed as measured solely from this crash. Full notes STAR: q-hs-crash.
They might ask next
If they ask
You answer
Pick one
IQVIA if they have not heard that intern. Ylogx if IQVIA already filled Invent/Learn/Backbone or the DOM story.
Horizon field “demo”
Field tests under ERC, not a SaaS client. 60 FPS operator feed versus stale occupancy. Different product.
40% and 99.9% are Ylogx product families. The crash is the defect class you fixed.
These two stories are notes. They are not a fourth primary on either intern if the LP spread is already full.
Tighten
One demo. If they already heard it on collab or Dive Deep, use the other product. Then stop.
Strengths
One sentence each: isolation (Ylogx), retrieval plus evals (IQVIA), date-driven systems (Horizon). Not sixteen Leadership Principles. Not a ranking of internships — three different jobs.
Class C opener. If they turn it into “not good at,” q-et-not-good. Other loop, not a forecast for Job 10454435.
What you say
Situation. They want what I actually ship, not a poster list.
Task. Three sentences. Three jobs. Then stop.
Action. Isolation: at Ylogx I put three-tier RLS in Postgres and the tenant in the Redis key so a chatbot cannot leak a JOIN. Retrieval and evals: at IQVIA I ship hybrid retrieval with cite-or-abstain and LangSmith so a fluent paragraph without a span does not count as done. Date-driven systems: at Horizon I sequenced camera and costmap software to an ERC date that does not move.
Result. Ylogx: −35% bot DB and 99.9% on that product. IQVIA: 200+ websites and 200+ page BRDs, no extra percent. Horizon: we 17th / 80+. Then I wait for which door they want.
They might ask next
If they ask
You answer
Give an example
The intern they point at. One module. Chapter 5 or the matching LP card.
Weakness / not good at
q-et-not-good. Honest intern gap. Not fake humility. Not “I work too hard.”
Name a Leadership Principle
Answer the principle on chapter 9. Do not recite all sixteen as strengths.
Keep straight
This is not Customer Obsession through Success and Scale as a speech.
Argus 73→89, GiftedBooks 3–10 min, StratifyLabs −30% wait until they pick those projects.
99.9% is Ylogx. IQVIA has no uptime percent.
Tighten
Three sentences. “Which do you want?” Stop. Why SDE I is the title card if they asked that instead.