11. Resume mock — they have the August 2026 PDF open
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 Round 4 only. Live Code is Java. Job 10454435 still has no public write-up that names a live-round question. Other people’s leftover hours grill the PDF. They are shape evidence, not a list for this Job ID. Unnamed stays unnamed. Click a stem, speak the card, stop after one number family.
Long run-on bullets — speak one job, then one number
The PDF packs two clauses into one dash. Awkward phrases like “AI chatbot using LangChain SQL RAG architecture” and “connected with RViz and Gazebo” are labels, not a speech. Speak one job-shaped sentence, then one number. If they want the second clause, they will ask.
What they might ask
Read this Ylogx chatbot bullet out loud. What did you ship?
What does “connected with RViz and Gazebo” actually mean?
Give me that line in twenty seconds.
Resume-derived wording on the August 2026 PDF. Other loop, not a forecast for Job 10454435.
What you say
Situation. At Ylogx, November 2024 to October 2025, the PDF line is one dash covering a chatbot, isolation, and Redis.
Task. Speak the job, then one number. Leave the rest for a follow-up.
Action.
I shipped a LangChain SQL RAG chatbot so people who were not writing SQL could ask warehouse questions.
Generated reads ran as the caller’s database role so three-tier RLS still applied.
Redis sat on repeat bot work with the tier in the key.
Result. Analysis productivity moved 65 percent. Stop. Isolation and Redis −35 percent wait for the next question.
They might ask next
If they ask
You answer
Clean that chatbot sentence again
I shipped LangChain SQL RAG so non-SQL users could query the warehouse behind three-tier RLS. Productivity +65%.
And Redis?
Bot database latency fell 35 percent. That is a second number family. See −35%.
Horizon “connected with RViz and Gazebo”
I took ZED 2 mapping at 2M+ points per second into RViz and Gazebo, and I put costmap planning with sensor fusion in front of the actuators. Obstacle +40%, collision −55%.
The title on every work row is intern. Architected still unpacks into modules you implemented. Say intern, then name the slice. Credit the team where the result is a team score.
What they might ask
You wrote Architected. You were an intern. Who actually owned this?
What did you design versus what already existed?
Did you lead a team?
Resume-derived IQVIA “Architected” twice. IE-asked leftover hours that probe intern vs lead. 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. The PDF uses Architected on the Deep Research line.
Task. Unpack that verb into the modules I shipped. Intern is still the title. I did not hire anyone.
Action.
I owned the LangGraph wiring: planner into typed state, then parallel tools — Firecrawl, Bing, DuckDuckGo, Google Playwright — so a dead scrape does not wipe surviving hits.
I owned the ranker versus first-hit default: quality, recency, and agreement of sources.
I owned the writer from retained graph state, and LangSmith traces so an empty scrape reaching the writer is visible.
Result. Ranking across 200+ websites. No IQVIA percent or latency on the resume.
They might ask next
If they ask
You answer
The second Architected line
Hybrid RAG: Azure AI Search Hybrid + Semantic, GraphDB hops, LangGraph choosing retrieval mode, LangSmith evals and test-case generation on 200+ page BRDs. Same intern, second system.
Ylogx Built / Designed
I owned report-builder path, SQL RAG behind RLS, Redis on the bot path, and dashboard deploy onto CloudFront / ECS / ALB. Intern title.
Horizon
Core software, not team lead. I owned GStreamer, ZED 2 mapping, costmap. 17th of 80+ is we. Mechanical and electrical are not mine.
Did you manage people?
No. IEDC Tech Team member. No mentee count on the resume.
What did you not own at IQVIA?
Building Azure AI Search itself. No IQVIA uptime percent.
Keep straight
200+ websites is Deep Research. 200+ page BRDs is Hybrid RAG. Two systems, one intern.
I for modules I shipped. We for ERC 17th / 80+, Argus site −50% / 2× after alerts, and team hackathon prizes.
Kubernetes / EKS cluster owner is not a title on these intern bullets. Ylogx orchestrator is ECS + Docker.
Tighten
Stop after the module list and 200+ websites. If they already heard Deep Research this hour, unpack Hybrid RAG modules, or switch to Ylogx Built/Designed on the report-builder path.
The PDF says ranking information for result generation. That is a mechanism: fan-out, then score quality, recency, and agreement. The number you defend is 200+ websites. There is no nDCG and no gold-set size on the resume.
What they might ask
What does “properly ranking information” mean?
What ranking metric did you use?
How large was the labeled set?
Resume-derived IQVIA Deep Research, 200+ websites. Other loop, not a forecast for Job 10454435.
What you say
Situation. On IQVIA Deep Research, a first search hit can look confident while two high-ranked pages contradict each other. A fluent writer will narrate a compromise neither page said.
Task. I needed ranking that treats disagreement as a failure mode, not a nicer paragraph.
Action.
I fan out Firecrawl, Bing, DuckDuckGo, and Playwright in parallel so one dead tool does not restart the run.
I rank sources on quality, recency, and agreement. Two credible pages that disagree is the bug.
I refuse a sentence without a retrieved span. LangSmith traces whether an empty scrape reached the writer.
Result. Research across 200+ websites. That coverage number is on the resume. nDCG is not.
No. Agreement of sources. 200+ websites is the resume number.
Gold-set size?
I do not have a labeled gold-set size on the resume.
Latency SLA?
IQVIA has no latency percent on the resume.
LangSmith scores ranking?
LangSmith traces evals. It does not print a public ranking study on this PDF.
Keep straight
200+ websites is Deep Research coverage. 200+ page BRDs is Hybrid RAG document size.
LangSmith is IQVIA. Ylogx chatbot is LangChain SQL RAG, not this ranker.
IQVIA has no uptime or extra percent on the resume. 99.9% is Ylogx.
Tighten
Stop after agreement plus 200+ websites. If they already heard Deep Research this hour, switch to Hybrid RAG on 200+ page BRDs (Invent), not a second ranking speech.
The mechanism is thirty live KPI dashboards instead of operators waiting on ad-hoc reports. The resume number is 60 percent. The PDF has no study protocol.
What they might ask
How did you get 60% operational efficiency?
Who did you survey? What was the window?
Is that the chatbot?
Resume-derived Ylogx 30 dashboards, +60%. Other loop, not a forecast for Job 10454435.
What you say
Situation. At Ylogx, operators waited on ad-hoc reports and one-off SQL for the same KPIs.
Task. I owned a live dashboard path they could open without a human in the middle of every grain.
Action.
I shipped 30 real-time KPI dashboards in React and Recharts.
Those GETs hit a deterministic serving path — CloudFront, ALB, ECS, Docker — with no LLM on that request.
I speak the resume number, then the evidence boundary.
Result. Operational efficiency improved 60 percent. I do not have the study protocol on the resume — not N, not the window, not who was surveyed.
No. Chatbot productivity is +65%. This family is dashboards.
Is 60% sub-210 ms?
No. sub-210 ms is dashboard latency. 60% is ops efficiency. Same product, two families.
CloudWatch as the source?
The PDF does not name a monitoring vendor as the source of 60%.
What would you add next?
A written baseline and protocol next to the tiles.
Keep straight
+60% is Ylogx dashboard ops. +65% is SQL RAG productivity. 40% is report-builder speed. Three Ylogx families.
sub-210 ms is the dashboard path with no LLM on that GET.
99.9% is Ylogx serving. GiftedBooks 99.5% is a different product.
Tighten
Mechanism, 60%, protocol sentence, stop. If they already heard Ylogx dashboards this hour, switch to Horizon GStreamer + costmap (Ownership backup), not a fourth Ylogx percent.
GiftedBooks +50% engagement — no student headcount
Engagement, reading, comprehension, and doubt time are one GiftedBooks family. The PDF has no student headcount and no survey instrument. GitHub currently describes AegisAI. Speak the PDF product.
What they might ask
How did you get 50% student engagement?
How many students? Randomized trial?
Walk me through GiftedBooks metrics.
Resume-derived GiftedBooks +50% / +35% / 2.5× / 3–10 min / sub-300 ms / 99.5%. Other loop, not a forecast for Job 10454435.
What you say
Situation. GiftedBooks is a VR learning suite: 3D labs, AI avatars, and PDF question-answering. A generic textbook voice does not answer from the student’s upload.
Task. Students ask from their PDF, get an answer on that material, and get PYQ-prioritized topics instead of a generic outline.
Action.
RAG on the student’s upload, grounded to retrieved spans.
API path held at sub-300 ms. Hosting number on this product is 99.5% uptime.
PYQ-prioritized topics sit beside the Q&A, not as a second chatbot personality.
Result. Engagement +50%. Reading efficiency +35%. Comprehension 2.5×. Doubts from hours to 3–10 minutes. I do not have the study protocol on the resume — not headcount, not a survey instrument.
They might ask next
If they ask
You answer
Student N?
Not on the resume. Protocol sentence, stop.
Is this AegisAI?
GitHub README currently describes AegisAI. I am answering the PDF product. README card.
LangSmith here?
LangSmith is IQVIA. GiftedBooks has no LangSmith claim.
sub-300 vs sub-210
sub-300 ms is GiftedBooks. sub-210 ms is Ylogx dashboards.
99.5% vs 99.9%
99.5% is GiftedBooks. 99.9% is Ylogx serving.
Keep straight
+50% engagement, +35% reading, 2.5× comprehension, hours → 3–10 min, sub-300 ms, and 99.5% are GiftedBooks.
Customer Obsession speaks 3–10 min first. Earn Trust can own the README mismatch if they opened GitHub.
Ylogx +60% ops is a different product.
Tighten
Mechanism, the GiftedBooks family, protocol sentence, stop. If they already heard GiftedBooks this hour, backup Customer with Ylogx SQL RAG +65%, or Earn Trust with Ylogx three-tier RLS.
40% faster reports — not sub-210 ms, not Redis −35%
Forty percent is the custom report-builder path. Sub-210 ms is dashboards. Redis −35% is bot database latency. Three numbers, one intern, three paths.
What they might ask
40% faster reports — faster than what?
Was that p95? Redis?
Is that the dashboard number?
Resume-derived Ylogx 40% reports. Other loop, not a forecast for Job 10454435.
What you say
Situation. At Ylogx, report generation waited on a human and one-off SQL for the same warehouse grain.
Task. I owned a custom report-builder path: FastAPI, NestJS, REST, PostgreSQL, so the warehouse query did not go through a person each time.
Action.
NestJS owned auth, RBAC, and report CRUD.
FastAPI and Postgres served the report query path as the caller’s role, same isolation idea as the bot.
I compare to the same product before that builder existed. I do not have the before/after query-log window on the resume.
Result. Reports 40% faster. That is this family. Sub-210 ms is dashboards. Redis −35% is the bot database path.
They might ask next
If they ask
You answer
Faster than 10 seconds?
The PDF has no timestamps. Mechanism plus 40%. Protocol sentence if they push N.
Did Redis cause 40%?
No. Redis −35% is bot database latency.
Is this sub-210 ms?
No. Dashboard path, no LLM on that GET.
p90 on the PDF?
p90 under 210 ms is written on the dashboard bullet, not on this 40%.
Client aggregation crash?
Heavy aggregation belonged off the browser. That defect class is on this product. 40% is not measured solely from that crash.
Keep straight
40% is report-builder speed. +65% is SQL RAG productivity. +60% is dashboard ops. −35% is bot DB via Redis. sub-210 ms is dashboards.
99.9% is Ylogx serving on ECS / ALB / CloudFront.
IQVIA has no report-speed percent.
Tighten
Stop after 40% and the split from dashboards and Redis. If they already heard Ylogx reports this hour, Deliver backup stays Horizon 17th / 80+ if the rover is still unused.
Sixty-five percent is people who were not writing SQL getting warehouse answers behind three-tier RLS. It is not “the LLM is 65% more accurate.”
What they might ask
65% data-analysis productivity — how do you know?
Is the model 65% more accurate?
Did the bot run as a superuser?
Resume-derived Ylogx SQL RAG +65%, three-tier RLS. Other loop, not a forecast for Job 10454435.
What you say
Situation. At Ylogx, people waited on an engineer for every grain of the warehouse. A chatbot that “worked” on a table-owner connection would still be a leak.
Task. Natural language to SQL that is usable and isolated.
Action.
LangChain SQL RAG: NL → constrained SQL over allow-listed schema, as the caller’s database role.
RLS and RBAC for three organizational tiers in Postgres, so a legal JOIN still lost the other tier’s rows.
I returned the SQL with the result so a human could see grain. A fast leak is not productivity.
Result. Analysis productivity +65%. I do not have the study protocol on the resume. I will not invent tickets closed per week. This is not model accuracy.
They might ask next
If they ask
You answer
Accuracy of the LLM?
+65% is analysis productivity for those users, not a model accuracy score.
Superuser SQL?
No. Caller role. RLS still applies.
LangSmith on this bullet?
LangSmith is IQVIA. This chatbot is LangChain SQL RAG.
LangGraph here?
LangGraph is IQVIA workflow state. Ylogx chatbot is LangChain.
Redis in this number?
Redis −35% is bot database latency, a neighboring family, not this 65%.
Keep straight
+65% is Ylogx SQL RAG productivity. Isolation is three-tier RLS. Redis −35% is bot DB.
Customer Obsession backup, if GiftedBooks already appeared, is this +65% story.
Literal tier names are not printed on the resume. Speak three organizational tiers.
Tighten
Stop after SQL RAG + RLS + 65%. If they already heard this chatbot this hour, do not add Redis as a fourth Ylogx trophy unless they ask Dive Deep / Frugality.
Redis is evidenced on the Ylogx bot path. Cache-aside, tier in the key, RLS still on a hit. No TTL and no p95 on the resume. This is not a rate-limiter product you shipped.
What they might ask
35% less bot database latency — Redis?
What TTL? What p95?
Did you ship a rate limiter?
Resume-derived Ylogx Redis −35%. Other loop, not a forecast for Job 10454435.
What you say
Situation. At Ylogx, every natural-language turn was hitting Postgres. It felt like “the model is thinking” when the database was the cost. Repeat questions and schema lookups are not proof the instance is undersized.
Task. Cut hot-path database work without weakening three-tier isolation.
Action.
I measured bot database latency, not dashboard GETs.
I put Redis cache-aside on repeat answers and schema lookups, with organizational tier and role in the key.
On a miss, SQL still ran as the caller role with RLS. If Redis was down, Postgres as that role — never table-owner, never fail-open isolation.
Result. Bot database latency −35%. I will not invent a TTL or a p95. A larger RDS was the spend I did not take first.
They might ask next
If they ask
You answer
Why still RLS after cache?
A cached answer for the wrong organization is still a leak.
TTL / p95?
Not on the resume.
Is this sub-210 ms?
No. sub-210 ms is dashboards, no LLM on that GET.
Rate limiter?
This Redis is cache on the bot path. A rate-limiter product is not an intern bullet. Independent SDE I live-round rate-limiter count is 1, not a UTA leftover default.
CloudFront for answers?
Private tenant JSON is not a public asset. CloudFront is static/dashboard edge.
Keep straight
−35% is Ylogx bot database latency via Redis. sub-210 ms is Ylogx dashboards.
Redis is evidenced here. Redis as a CDN, as RLS, or as a shipped rate limiter is a different claim.
Ylogx primaries for this number: Frugality and Dive Deep (with RLS). Not a fourth Invent trophy.
Tighten
Stop after −35% with isolation still on. If they already heard Ylogx Redis this hour, Frugality backup is Horizon fusion versus extra hardware.
Ninety-nine point nine is the Ylogx serving path: ECS desired count, ALB health checks, CloudFront, CI/CD. The PDF has no monitoring-vendor screenshot. GiftedBooks uptime is 99.5%.
What they might ask
99.9% uptime — show me the dashboard.
Which vendor? Which SLO window?
Is IQVIA also 99.9%?
Resume-derived Ylogx 99.9%; GiftedBooks 99.5%. Other loop, not a forecast for Job 10454435.
What you say
Situation. At Ylogx the serving path had to stay up for reports, the bot, and dashboards. Patching a snowflake host is not that path.
Task. Speak the resume number from mechanisms I owned, not from a vendor board I did not put on the PDF.
Action.
Dockerized services on ECS with a desired count of tasks.
ALB in front with health checks so a sick task is not left in rotation. Fail a check → ALB stops sending traffic; ECS replaces the task.
CloudFront at the edge for the static/dashboard path. CI/CD so we were not SSHing a snowflake host.
Result. 99.9% uptime on that Ylogx serving path. I do not have the incident CSV or SLO window on the resume.
They might ask next
If they ask
You answer
Datadog / New Relic / PagerDuty?
The PDF has no monitoring-vendor name. I describe ECS, ALB, CloudFront, CI/CD.
Kubernetes?
Ylogx orchestrator on the intern bullets is ECS + Docker. See Kubernetes card.
GiftedBooks 99.5%
Different product, different hosting number. Not this 99.9%.
IQVIA uptime?
IQVIA has no uptime percent on the resume.
Missed 99.9%?
Not on the resume. Failure story on this PDF is Argus 73% mAP as an unshippable alert bar.
Keep straight
99.9% uptime is Ylogx. 99.5% is GiftedBooks. IQVIA has no uptime percent.
Nginx is on the skills line. Ylogx edge on the intern bullets is CloudFront → ALB → ECS.
Vercel / Render / Railway are tools-line names. Ylogx production hop is CloudFront / ECS / ALB.
Tighten
Stop after ECS / ALB / CloudFront and 99.9%. If they already heard Ylogx serving this hour, Ownership backup is Horizon GStreamer + costmap.
Horizon +40% obstacle / −55% collision — rover, not Argus
These percents sit on GStreamer, ZED 2, and costmap fusion. They are not Argus mAP. 17th of 80+ is we. Core software, not team lead.
What they might ask
Horizon 40% obstacle detection and 55% collision risk — protocol?
Is that YOLO mAP on the rover?
Did you design the chassis?
Resume-derived Horizon +40% / −55%, 60 FPS, 2M+ pts/s, 17th / 80+. Other loop, not a forecast for Job 10454435.
What you say
Situation. February to June 2024 I was core software on Team Horizon’s ROS2 rover under an ERC date that does not move. A live operator feed with a stale costmap is two systems, one of them failing.
Task. Own the software path that had to run on the day: camera, mapping, costmap in front of the actuators.
Action.
GStreamer camera pipeline at 60 FPS so the operator saw the newest frame.
ZED 2 mapping at 2M+ points per second into occupancy, with a separate branch into RViz and Gazebo so visualization did not starve planning.
Costmap fusion on the sensors we had, not a purchase order for another camera as the first fix.
Result. Obstacle detection +40%. Collision risk −55%. We placed 17th of 80+ at ERC 2024. I do not have the study protocol on the resume — not a labeled bag count.
They might ask next
If they ask
You answer
Is +40% Argus mAP?
No. Argus is 73% → 89% mAP on PPE. This is rover obstacle / collision.
60 FPS vs 24 FPS
60 is Horizon GStreamer. 24 is Argus detector budget.
Did you design the rocker-bogie?
No. Mechanical and electrical are not mine. I owned software.
Missed ERC?
Not on the resume. We placed 17th / 80+.
Team lead?
Core software member. 17th is we.
Keep straight
+40% obstacle and −55% collision are Horizon. 73% → 89% mAP is Argus.
60 FPS is rover camera. 24 FPS is PPE detector. StratifyLabs −30% is ML iteration time, not FPS.
Horizon is a competition rover internship, not a SaaS internship.
Tighten
Stop after +40% / −55% and 17th as we. If they already heard Horizon this hour, Deliver backup is Ylogx 40% reports + 99.9%; Bias backup is eight named hackathons.
Detector eval is 73% → 89% mAP on 15,000+ images. Site outcomes are −50% violations and 2× compliance after humans got alerts they could trust. 73% is not shippable for floor alerts.
What they might ask
50% fewer violations and 2× compliance — is that mAP?
So 89% of workers wear PPE?
When did you scale to 20+ cameras?
Resume-derived Argus 73→89% mAP, 24 FPS, 20+ cameras, −50% / 2×. Other loop, not a forecast for Job 10454435.
What you say
Situation. Argus is PPE computer vision. At 73% mAP you miss helmets and you fire junk boxes. Either one trains people to ignore alarms.
Task. Treat 73% as a fail. Ship a detector a human would page on, then scale cameras.
Action.
I trained YOLOv9 on 15,000+ images and held an eval gate: 89% mAP and 24 FPS before 20+ cameras.
Alerts logged to Postgres. OpenCV decodes and resizes; YOLOv9 is the detector.
Only after that gate did the site run 20+ containerized feeds.
Result. 73% → 89% mAP and 24 FPS are detector numbers. We saw −50% violations and 2× compliance after alerts existed. I do not have the site study protocol on the resume — not a plant name, not a shift window.
They might ask next
If they ask
You answer
Is 89% mAP 89% of pixels?
No. mAP is mean average precision over PPE classes after IoU matching. Not accuracy, not “89% of people.”
Scale at 73%?
No. Gate is 89% mAP and 24 FPS, then 20+ cameras.
TensorFlow as this jump?
TensorFlow is listed. Quoted jump is YOLOv9 73→89% mAP.
Horizon 24 FPS?
Horizon camera is 60 FPS GStreamer. Argus budget is 24 FPS.
LLM as the detector?
YOLOv9. Captions are not this detector.
Keep straight
73% → 89% mAP and 24 FPS are detector eval. −50% / 2× are site outcomes after alerts. 20+ cameras come after the gate.
I for training, eval gate, containerized feeds. We for site −50% / 2× after alerts.
Argus primaries: Are Right, Highest Standards, Success and Scale. Horizon +40% / −55% stay on the rover.
Tighten
Stop after the split: mAP is not the site percent. If they already heard Argus this hour, Are Right backup is Ylogx measured bot-DB −35%.
Most resume percents have a mechanism and a number, and no methods appendix. Name what you shipped, say the PDF number, then: “I do not have the study protocol on the resume.” That is the honest bar, not a shrug.
What they might ask
How did you measure that?
What was N? What was the window?
Walk me through the study.
IE-asked leftover hours that grill every metric (LC 7981646 shape). Resume-derived percentages without a protocol on the PDF. Other loop, not a forecast for Job 10454435.
What you say
Situation. They point at Ylogx 60% operational efficiency — or any percent without a methods line.
Task. Answer with the mechanism I owned, the exact resume number, then the evidence boundary.
Action.
Name the shipped path: thirty React/Recharts dashboards instead of ad-hoc report waits.
Say the resume number: 60 percent.
Say: I do not have the study protocol on the resume. I will not invent N, a window, or who was surveyed.
Result. They still have a real system. They do not have a fake study. Adjacent numbers I will defend on the same product if they stay: 40% reports, sub-210 ms dashboards, 99.9% uptime, Redis −35% on the bot path — each as its own family.
Java on skills / Live Code vs intern Python, TypeScript, ROS2
This loop’s editor is Java. Internship bullets used Python FastAPI, TypeScript NestJS/React, and ROS2. Argus YOLOv9 is Python. Same ideas — maps, heaps, graphs, isolation — transfer. The intern products did not run on Spring.
What they might ask
Why is Java on the resume if internships were Python?
Can you write this in Java?
Did IQVIA run on Spring Boot?
Resume-derived skills line Java. Live Code Java. Intern bullets Python / TypeScript / ROS2. Other loop, not a forecast for Job 10454435.
What you say
Situation. They see Java on the skills line and Python FastAPI on IQVIA.
Task. Say the split once, honestly, then write Java in the editor if they paste a problem.
Action.
Java is a language I write for this Live Code loop, and it is on the skills line.
Production intern work was Python FastAPI, TypeScript NestJS and React, ROS2 on Horizon, YOLOv9 in Python on Argus.
If this shop wants the same agent graph in Java services, the ideas transfer: nodes, state, retrieval interfaces. Syntax is not the architecture.
Result. I will not pretend IQVIA ran on Spring. I will write Java in Live Code.
They might ask next
If they ask
You answer
Spring / Kafka / EKS on internships?
Those names are not on the intern bullets. Internships used Python, TypeScript, and ROS2. Live Code is Java.
TypeScript where?
Ylogx NestJS (auth, CRUD) plus React dashboards. StratifyLabs is a Next.js product on the resume.
They paste code after an LP-only invite
Write Java. Other loops were told LP-only and still got DSA. Shape evidence, not a forecast.
HashMap vs Python dict
Same idea. I implement in Java in the editor.
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.
Argus detector is YOLOv9 in Python. Horizon is ROS2 / GStreamer / ZED.
Say this split once. Later answers name the module, not the language debate.
Tighten
One split, then stop. If they already heard this, go to the intern module they named — IQVIA retrieval, Ylogx RLS, or Horizon costmap — in the language of that bullet.
Kubernetes is on the skills line. The orchestrator on the Ylogx intern bullets is ECS + Docker. You are not an EKS cluster owner on this resume.
What they might ask
You listed Kubernetes. Where did you run it in production?
Was 99.9% on EKS?
Pods versus tasks?
Resume-derived skills Kubernetes; Ylogx CloudFront / ECS / Docker / ALB. Other loop, not a forecast for Job 10454435.
What you say
Situation. At Ylogx I deployed the BI serving path on AWS.
Task. Speak the hop I will defend, not the skills-line word they circled.
Action.
A task definition is the recipe: image, CPU, memory, ports. A task is one running copy. A service keeps a desired count and attaches the ALB target group.
Health checks drop a sick task from rotation; ECS starts a replacement. That is how I talk about 99.9% without a vendor screenshot.
Docker images are the unit. Kubernetes knowledge is listed. This intern product is ECS + Docker.
Result. 99.9% on that ECS / ALB / CloudFront path. I will not invent Fargate versus EC2 launch type, vCPU, or an EKS cluster I operated.
They might ask next
If they ask
You answer
Task vs service
Recipe versus desired count of running copies behind the ALB.
ALB vs CloudFront
CloudFront is edge/CDN. ALB is regional L7 with health checks. Client → CloudFront → ALB → ECS.
Nginx as the edge?
Nginx is listed. Ylogx edge on the bullets is CloudFront → ALB → ECS.
Argus containers
Argus is containerized 20+ feeds after the mAP gate. Still not an EKS-owner claim.
Keep straight
Kubernetes is listed. Ylogx intern orchestrator is ECS + Docker. EKS cluster owner is not a title on this resume.
99.9% is Ylogx serving. GiftedBooks 99.5% is a different product.
GoDaddy is the registrar hop into Route 53. Route 53 is the zone. They are not the same sentence as “I ran EC2 by hand.”
Tighten
ECS + ALB + CloudFront, 99.9%, stop. If they already heard Ylogx deploy this hour, Success and Scale backup is Argus 20+ cameras after 89% / 24 FPS.
Listed means you have used the tool enough to name it. Evidenced means it sits on an intern or project bullet with a resume number. Redis is evidenced on the Ylogx bot path. Redis as a rate-limiter product is not. Point to chapter 8 when they keep reading the skills block like a menu.
What they might ask
GraphQL, ProtoBuf, Nginx, Auth0 — where in production?
Is FAISS the IQVIA index?
You listed Redis. Did you ship a rate limiter?
Resume-derived skills block versus intern bullets. Other loop, not a forecast for Job 10454435.
What you say
Situation. They read the skills block as if every name were a Ylogx production hop.
Task. One honest sentence, then the evidenced hop they actually asked about.
Action. Listed means I have used the tool enough to name it. I will not fake production on a bullet that does not exist. Redis is evidenced: Ylogx bot-database cache, −35%, tier in the key, RLS still on. That Redis is not a rate-limiter product I shipped. Independent SDE I live-round rate-limiter count is 1, not a UTA leftover default.
Result. They get one evidenced path. The rest of the menu stays a lookup table, not a tour.
They might ask next
Skill
Where it lives
What you speak
GraphQL
Listed
Ylogx intern path is REST.
ProtoBuf
Listed
Not on an intern bullet. Horizon live video is a GStreamer pipeline, not GraphQL.
Nginx
Listed
Ylogx edge is CloudFront → ALB → ECS.
OAuth / JWT
Skills
Ylogx isolation is RBAC at NestJS + RLS at Postgres. JWT as an API identity pattern is honest CS.
Auth0
Listed
Auth0 as Ylogx production is not the named hop. RLS is Postgres, not Auth0.
Redux
Listed
Dashboards say React.js and Recharts.
Prisma
Listed
Warehouse bullet is PostgreSQL.
Mongo / MySQL / Firebase / Supabase
Listed
Source of truth I will defend is PostgreSQL (Ylogx; Argus event log).
Redis
Evidenced
Ylogx −35% bot DB. Cache-aside, tier keys. Not a rate limiter you shipped.
TensorFlow
Listed
Argus quoted jump is YOLOv9 + OpenCV, 73→89% mAP.
FAISS
Listed as Vector DB
IQVIA intern retrieval is Azure AI Search Hybrid + Semantic. GiftedBooks is the student-PDF lane.
Neo4j vs GraphDB
Skills say Neo4j; IQVIA bullet says GraphDB
I say GraphDB. Neo4j if they point at the skills line.
OpenAPI
Listed (Postman, OpenAPI Spec)
How I describe REST contracts. Not a billed API-gateway product.
WebSockets
Listed
30 KPIs: live-ish REST/poll unless they ask. I will not invent Redis pub/sub fan-out for all 30 tiles.
Redis −35% is Ylogx bot database cache. A rate-limiter product is not on the intern bullets. Independent SDE I rate-limiter count is 1, not UTA-default Round 4.
FAISS is listed. IQVIA retrieval is Azure AI Search + GraphDB.
LangSmith is IQVIA; Ylogx chatbot is LangChain SQL RAG
LangSmith traces evals and test-case generation on Deep Research and Hybrid RAG. Ylogx chatbot is LangChain SQL RAG, not LangGraph. Do not retrofit traces onto the BI bot.
What they might ask
Is LangSmith on the Ylogx chatbot?
Is all your RAG LangGraph?
Where are the traces?
Resume-derived IQVIA LangSmith; Ylogx LangChain SQL RAG. Other loop, not a forecast for Job 10454435.
What you say
Situation. I am a software intern at IQVIA. Two FastAPI systems share LangGraph and LangSmith. They fail differently.
Task. Traces have to show whether the ranker, an empty scrape, or the writer failed — including generated tests on BRDs.
Action.
Deep Research: LangSmith on node order so an empty scrape reaching the writer is visible. Ranker is agreement, not first-hit.
Hybrid RAG: LangSmith gold items carry expected citation IDs; test-case generation is itself a traced job. Cite-or-abstain.
I do not log full confidential BRDs into traces in production.
Result. Evals and test-case generation with LangSmith on IQVIA. No extra IQVIA percent.
They might ask next
If they ask
You answer
Ylogx chatbot
LangChain SQL RAG behind three-tier RLS. Redis −35% on that bot path. No LangSmith on that bullet.
LangGraph on Ylogx?
LangGraph is IQVIA workflow state. Ylogx bot is LangChain.
GiftedBooks traces?
GiftedBooks is student PDF RAG. No LangSmith claim.
GraphDB vs LangGraph
GraphDB stores requirement dependencies. LangGraph is agent workflow state.
Why two IQVIA systems?
Deep Research has no single gold page — 200+ websites. Hybrid RAG has a gold document — 200+ pages. Two retrieval backends, one orchestrator family.
Keep straight
LangSmith is IQVIA. Ylogx chatbot is LangChain SQL RAG, not LangGraph.
200+ websites vs 200+ page BRDs are two IQVIA families.
GiftedBooks and StratifyLabs do not inherit LangSmith.
Tighten
Stop after IQVIA traces. If they already heard IQVIA this hour, do not steal LangSmith onto Ylogx — switch to Ylogx RLS if they asked isolation.
Resume wins. Label the mismatch in the first sentence, then speak GiftedBooks from the August 2026 PDF. There is no merge timeline on the resume to narrate.
What they might ask
I opened github.com/adarshx01/giftedbooks. This README is AegisAI.
Are they the same product?
Resume-derived GiftedBooks PDF vs current README. Laptop-open hygiene. Other loop, not a forecast for Job 10454435.
What you say
Situation. They opened the GiftedBooks repo. The README currently describes AegisAI.
Task. Name the mismatch. Walk the PDF product. Offer to fix the README.
Action.
That is a mismatch. I am answering the August 2026 resume, not the wrong README.
GiftedBooks is a VR learning suite: 3D labs, AI avatars, RAG on the student’s PDF, PYQ topics.
I would fix the README. I will not mix AegisAI features into this story.
Resume wins. Speak stratifylabs.design: browser inference, −30% ML iteration, marketplace 50+, URDF/WebGL, Gemini in-scene bots. The default README is not the product.
What they might ask
StratifyLabs GitHub is a default Next.js README.
Is −30% camera FPS?
Is the marketplace 50 GPUs?
Resume-derived StratifyLabs −30% / 50+. Other loop, not a forecast for Job 10454435.
What you say
Situation. They opened StratifyLabs on GitHub. The README is a Next.js scaffold.
Task. Name that. Speak the live product on the resume.
Action.
The default README is not the product. I am walking stratifylabs.design.
Browser-based inference so iteration does not wait on a heavy local loop. 3D simulation lab, URDF editor with WebGL.
Marketplace catalog of models and datasets. Gemini RAG bots as 3D characters on the current sim.
Result. ML iteration time −30%. Marketplace 50+. −30% is iteration time, not FPS.
They might ask next
If they ask
You answer
Is −30% Argus 24 FPS?
No. Stratify −30% is ML iteration. Argus 24 FPS is detector budget. Horizon 60 FPS is GStreamer.
50+ GPUs?
Marketplace catalog 50+ models/datasets, not a GPU farm.
Deep RLS here?
Three-tier RLS is Ylogx.
LangSmith here?
LangSmith is IQVIA.
Horizon Gazebo?
Horizon RViz/Gazebo is rover software. Stratify WebGL is the browser lab.
If they open GitHub
Speak stratifylabs.design and the resume numbers. Features that exist only on the scaffold README stay off this answer.
Keep straight
StratifyLabs product is stratifylabs.design and the resume. The GitHub README is a Next.js scaffold.
−30% is ML iteration. 50+ is marketplace catalog. Think Big primary. Invent backup if IQVIA already appeared.
IQVIA 200+ sites is Think Big backup only, not a fourth Stratify trophy.
Tighten
Scaffold, then −30% and 50+, stop. If they already heard StratifyLabs this hour, Invent stays IQVIA Hybrid RAG if unused; Think Big backup is IQVIA 200+ sites.
valAgent / warpflow / karyanode are adjacent, not intern metrics
GitHub-adjacent experiments do not inherit IQVIA 200+, Ylogx 99.9%, or any other intern family. Login Tracker is unverified. Do not volunteer it as this Job ID.
What they might ask
What is valAgent?
Is warpflow the Ylogx product?
What is karyanode?
Resume-derived GitHub-adjacent vs intern bullets. Other loop, not a forecast for Job 10454435.
What you say
Situation. They opened valAgent, an ETL-validator on GitHub, while asking about IQVIA.
Task. Label adjacent. Do not merge 200+ websites onto a side repo.
Action. valAgent is GitHub-adjacent to research habits. It is not an IQVIA resume bullet. Deep Research on the PDF is LangGraph, parallel tools, agreement ranking, 200+ websites, LangSmith traces.
Result. One intern number family, on the intern product. Adjacent repo gets no extra percent.
You had three internships at three times. Each job is a different product. If they ask which intern ran longest in production, answer with a date: Ylogx, November 2024 to October 2025.
What they might ask
Walk me through your internships.
Why three internships?
Which intern ran longest in production?
IE-asked current role / intern walk (GFG 2025 leftover hour, other loop). Resume-derived dates. Other loop, not a forecast for Job 10454435.
What you say
Situation. They asked for the three internships as a walk, not for a winner.
Task. Speak dates and what each job was. Longest in production is a date if they ask it.
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, three times. If they ask which intern ran longest in production: Ylogx, November 2024 to October 2025.
They might ask next
If they ask
You answer
Longest in production?
Ylogx, November 2024 to October 2025. That is a date.
Current role?
IQVIA intern, Kochi, two FastAPI systems, then stop. Offer one door.
Why three?
Each has a different failure mode. Horizon: stale costmap versus a live 60 FPS feed. Ylogx: latency and isolation. IQVIA: first-hit ranking and vector-only BRDs. I would rather go deep on one than tour all three in ninety seconds.
They pick one door
Stay on that intern. Do not add StratifyLabs, GiftedBooks, and Argus onto the timeline unless they ask for projects.
Keep straight
Horizon Feb–Jun 2024 · Ylogx Nov 2024–Oct 2025 · IQVIA Apr 2026–present. Dates do not overlap each other.
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 extra percent.
Tighten
Three jobs, then one door. Stop. If they already heard the timeline, go deep on the intern they named — do not recast the same three as a contest.
Ylogx ends October 2025. IQVIA starts April 2026. Horizon is February–June 2024. School is October 2022–May 2026. Current role is IQVIA, April 2026–present, final-year intern in Kochi. The intern work is real. You do not invent a day-by-day calendar.
What they might ask
Education dates overlap IQVIA. How many hours?
Do Ylogx and IQVIA overlap?
Will you relocate to Karnataka?
Resume-derived date rows. Other loop, not a forecast for Job 10454435.
What you say
Situation. They see CUSAT through May 2026 and IQVIA from April 2026.
Task. Say the overlap is final-year internship, not a fake graduation date, and not a fake hour log.
Action. Final-year CSE with an intern role in Kochi from April 2026. Graduation window on the PDF is May 2026. Ylogx ended October 2025 before IQVIA started. Horizon had already ended June 2024. I will not invent a day-by-day calendar. The intern work is real: two FastAPI systems at IQVIA.
Result. Current role: Deep Research across 200+ websites and Hybrid RAG on 200+ page BRDs. No IQVIA percent.
They might ask next
If they ask
You answer
Ylogx vs IQVIA overlap?
They do not overlap. Ylogx ends October 2025. IQVIA starts April 2026.
Gap Nov 2025–Mar 2026?
I will not fill it with off-resume stories unless I have confirmed I want them on the record.
CGPA inflated?
CUSAT B.Tech CSE, 8.42 / 10. I will not round it up. It is education, not an LP metric.
Relocate to Karnataka?
I intern in Kochi and I studied at CUSAT. This role is ADCI Karnataka. I will relocate.
Already graduated in March?
Graduation window on the PDF is May 2026.
Keep straight
IQVIA April 2026–present is current. Ylogx November 2024–October 2025. Horizon February–June 2024.
CUSAT October 2022–May 2026, CGPA 8.42.
Django portal, Purplle CCTV, abliteration are prep-only unless you confirm them onto the record.
Tighten
Dates, current role, stop. If they already heard IQVIA as current, offer Ylogx isolation or Horizon costmap as the next door — not a hours spreadsheet.
Name three modules, name what you did not own, stop. If they already heard Ylogx this hour, unpack Horizon software vs hardware, or IQVIA two systems vs Azure itself.
One job-shaped sentence, then one number. If they want the second clause, they will ask. I for modules you shipped. We for ERC 17th / 80+, Argus site outcomes after alerts, and team hackathon prizes.
What they might ask
Give me that bullet in twenty seconds.
Walk the resume without dumping every percent.
Resume-derived August 2026 PDF lines. Other loop, not a forecast for Job 10454435.
What you say
Situation. They point at one PDF line.
Task. Speak the matching row below. Stop after the number in the last column.
Action. One sentence that is a job, not a stack dump. Then the number you will defend.
Result. They choose the follow-up. You do not pre-load the second family.
Spoken lines
Line
You say (~20s)
Stop on
IQVIA Deep Research
I am a software intern at IQVIA in Kochi, April 2026 to now. I built a LangGraph Deep Research service on FastAPI: a planner fans work to Firecrawl, Bing, DuckDuckGo, and Google Playwright; a ranker scores quality, recency, and agreement; a writer synthesizes from graph state so a dead scrape does not restart the run.
200+ websites
IQVIA Hybrid RAG
The second system is Hybrid RAG on 200+ page BRDs and PDFs: Azure AI Search Hybrid plus Semantic, a GraphDB for requirement hops, LangGraph choosing retrieval mode, and LangSmith traces for evals and generated test cases.
200+ page BRDs (no extra %)
Ylogx reports + uptime
At Ylogx, November 2024 to October 2025, I shipped a full-stack AI BI report builder: Python FastAPI, NestJS, REST, PostgreSQL. Reports got 40% faster. That serving path is the 99.9% uptime number.
40% reports · 99.9% if they ask serving
Ylogx SQL RAG + RLS + Redis
I shipped a LangChain SQL RAG chatbot so people who were not writing SQL could ask questions. I put RLS and RBAC in Postgres for three organizational tiers, and Redis on the bot path.
+65% · −35% bot DB as the second ask
Ylogx dashboards + AWS
I shipped 30 real-time KPI dashboards in React and Recharts. Deploy was CloudFront, ECS, Docker, CI/CD, GoDaddy DNS into Route 53 into an ALB. Dashboards held sub-210 ms. Resume says 60% operational efficiency.
30 dashboards · +60% · sub-210 ms
Horizon camera
February to June 2024 I was core software on Team Horizon’s ROS2 Mars rover. I owned the live camera path at 60 FPS with GStreamer. We placed 17th globally of 80+ teams at ERC 2024.
60 FPS · 17th / 80+ is we
Horizon costmap
I took ZED 2 mapping at 2M+ points per second into RViz and Gazebo and put costmap planning with sensor fusion in front of the actuators.
+40% obstacle · −55% collision
Argus
I built a real-time PPE detector with YOLOv9, 73% to 89% mAP on 15,000+ images, 24 FPS, then 20+ containerized camera feeds with Postgres logs.
89% mAP + 24 FPS · −50% / 2× after alerts
GiftedBooks
GiftedBooks is a VR learning suite with 3D labs, AI avatars, and RAG on the student’s PDF. API under 300 ms. Doubts from hours to 3–10 minutes.
3–10 min · +50% / +35% / 2.5× · 99.5%
StratifyLabs
StratifyLabs is a browser computer-vision lab plus marketplace: inference in the browser, URDF/WebGL, Gemini scene bots. Live site stratifylabs.design.
−30% ML iteration · marketplace 50+
IEDC / hackathons
I was Tech Team member of IEDC CUSAT, not a manager. Across eight national and regional hackathons we took first at CodeRecet, MLH Best Project, and Magnathon 2.0 runner-up. CUSAT CSE CGPA 8.42.
member · eight events · named prizes · 8.42
They might ask next
If they ask
You answer
Second clause on the same dash
Give the neighboring family on that row. Do not jump internships.
A leftover hour often asks why a metric exists. Mixing two products in one sentence is how the story falls apart. This table is a lookup, not a speech.
What they might ask
Is this the same 60 FPS as Argus?
Is 99.9% GiftedBooks?
Is −30% a camera rate?
Resume-derived August 2026 PDF families. Same map as Home.
What you say
Situation. They mixed Horizon 60 FPS with Argus 24 FPS.
Task. Name the product first, then the family. The table below is the lookup.
Action. 60 FPS is the rover camera pipeline I owned on Horizon with GStreamer. Argus 24 FPS is the PPE detector budget after the 89% mAP gate — a different product.
Result. One correction, then I stay on the product they meant.
They might ask next
Keep apart
Why
Horizon 60 FPS and Argus 24 FPS
60 is the rover camera pipeline. 24 is the PPE detector budget.
StratifyLabs −30%
ML iteration time in the browser. Not a camera rate.
Argus mAP and −50% / 2×
mAP is held-out detector quality. Violations and compliance are after alerts existed.
Horizon +40% / −55%
Rover obstacle and collision. Not Argus.
Ylogx −35% and sub-210 ms
−35% is bot database latency via Redis. sub-210 ms is dashboard serving with no LLM on that GET.
Ylogx +65%, 40%, +60%
SQL RAG productivity, report-builder speed, dashboard ops. Three families.
Ylogx 99.9% and GiftedBooks 99.5%
Different products.
GiftedBooks sub-300 ms and Ylogx sub-210 ms
Different APIs.
IQVIA 200+ sites and 200+ page BRDs
Deep Research coverage versus Hybrid RAG document size.
GraphDB and LangGraph
GraphDB stores requirement dependencies. LangGraph is workflow state.
73% mAP
Not a shippable alert. Gate is 89% mAP and 24 FPS, then 20+ cameras.
Keep straight
IQVIA has no uptime or latency percent. LangSmith is IQVIA. Ylogx chatbot is LangChain SQL RAG.
Redis −35% is Ylogx bot cache. Rate-limiter count as a shipped intern product is not on this PDF. Independent SDE I live-round rate-limiter count is 1.
Job 10454435 still has no public named live-round question. Unnamed stays unnamed.
Tighten
Lookup, correct the mix, stop. Then answer on the one product they meant.