# Resume mock — Bar Raiser / HM (Adarsh Vishwakarma)

Notes for Adarsh. Job **10454435** still has **no** public IE that names a live-round question. Other SDE I / UTA / AUTA later rounds grill the **PDF**, not a second victory lap. Live Code is **Java**. Internships shipped **Python / TypeScript** (ROS2 on Horizon; Argus YOLOv9 in Python). GiftedBooks answers = **Aug 2026 resume only**. GitHub `giftedbooks` currently describes **AegisAI** — mismatch. InstaRecon if they open GitHub: ethics one-liner, then StratifyLabs / Argus / Ylogx / IQVIA. Do **not** invent measurement protocols, N, windows, Datadog, Kafka, Qdrant, Auth0-as-Ylogx-prod, or EKS ownership.

**How to use.** They ask from the PDF. You answer in spoken sentences. After the resume number, stop. If they push “how did you measure,” say the **mechanism** you actually owned, then: *I do not have the study protocol on the resume.* That sentence is not humility theatre. It is the honest bar.

Format below: **Q** (they ask) → **Spoken** → **Trap**.

**I vs we.** **I** for the slice you owned. **We** for ERC 17th / 80+ (team score) and for hardware you did not design. Resume verbs *Architected / Built / Designed* still sit on an intern title — say intern, then name the module.

---

## 0. Twenty-second intern bullets (run-ons, cleaned)

The PDF packs two clauses into one dash. Speak **one job-shaped sentence**, then one number. If they want the second clause, they will ask.

### IQVIA — Deep Research

**Spoken (~20s).** 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, and a writer synthesizes from graph state so a dead scrape does not restart the run. The number I will defend is ranking across **200+** websites.

**Trap.** “Properly ranking information” is not a metric. Do not invent nDCG, a gold-set size, or a latency SLA. Do not call this LangSmith-on-Ylogx. Do not say Kubernetes.

### IQVIA — Hybrid RAG

**Spoken (~20s).** 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. Production is Python. This loop’s Live Code is Java.

**Trap.** Do not say Qdrant. Do not steal Redis −35% onto IQVIA. FAISS is on the skills line; this intern retrieval is Azure AI Search. GraphDB on the bullet; Neo4j only if they point at the skills line.

### Ylogx — report builder + uptime

**Spoken (~20s).** 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 on the resume.

**Trap.** Do not invent a Datadog board, an SLO window, or “three nines for twelve months.” Do not put an LLM on the dashboard hot path. **sub-210 ms** is the dashboard number, not the chatbot.

### Ylogx — SQL RAG + RLS + Redis

**Spoken (~20s).** 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. Analysis productivity moved **65%**. Bot database latency dropped **35%**.

**Trap.** Redis **is** evidenced here. LangSmith is **not**. Do not say the bot is a superuser. Do not name the three tiers (org/team/user) unless you confirm off-resume.

### Ylogx — dashboards + AWS

**Spoken (~20s).** I shipped **30** real-time KPI dashboards in React and Recharts. Resume says **60%** operational efficiency. Deploy was CloudFront, ECS, Docker, CI/CD, GoDaddy DNS into Route 53 into an ALB. Dashboards held **sub-210 ms**.

**Trap.** 60% has no protocol on the PDF. Speak mechanism (live KPIs instead of ad-hoc report waits), then the resume number, then *I do not have the study protocol on the resume.* Do not claim Nginx as the Ylogx edge. Do not claim Kubernetes.

### Team Horizon — rover + camera

**Spoken (~20s).** February to June 2024 I was **core software** on Team Horizon’s ROS2 Mars rover. We placed **17th** globally of **80+** teams at ERC 2024. I owned the live camera path at **60 FPS** with GStreamer.

**Trap.** 17th is a **team** score. Credit mechanical / electrical / ERC hardware. Do not say you designed the chassis. mAP belongs to Argus, not the rover.

### Team Horizon — mapping + costmap

**Spoken (~20s).** 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. Obstacle detection improved **40%**. Collision risk dropped **55%**.

**Trap.** Do not invent a missed ERC. Do not merge Argus YOLO onto the rover. Extra sensors were not the first spend.

---

## 1. Imprecise claims (mechanism + number + no protocol)

### Q. What does “properly ranking information” mean?

**Spoken.** It is not a published ranking metric. On Deep Research I do not take the first search hit as truth. I fan out Firecrawl, Bing, DuckDuckGo, and Playwright, then I score sources on quality, recency, and **agreement**. Two high-ranked pages that contradict each other is the failure mode: first-hit retrieval looks confident and the writer will narrate a fluent compromise that neither page said. I refuse a sentence without a retrieved span. The resume number is **200+** websites, not an nDCG. I do not have a labeled gold-set size on the resume.

**Trap.** “We used a state-of-the-art ranker.” “I A/B tested ranking.” Invented BLEU / RAGAS scores. LangSmith **traces** evals; it does not magically produce a public study.

### Q. How did you get 60% operational efficiency?

**Spoken.** The mechanism I owned is **30** live KPI dashboards on React and Recharts instead of operators waiting on ad-hoc reports and one-off SQL. The resume number is **60%**. I do not have the study protocol on the resume — not N, not the window, not who was surveyed. I will not invent them. Adjacent numbers I *will* defend on the same product: **40%** faster reports, **sub-210 ms** on that dashboard path, **99.9%** uptime, Redis **−35%** on the bot database path.

**Trap.** “I ran a six-month ops study.” Fake baseline hours. Putting 60% on the chatbot. Claiming CloudWatch as the source of 60%.

### Q. How did you get 50% student engagement on GiftedBooks?

**Spoken.** GiftedBooks is a VR learning suite with 3D labs, AI avatars, and PDF Q&A. The mechanism is: students ask from **their** upload, get an answer in **sub-300 ms**, and get PYQ-prioritized topics instead of a generic textbook voice. The resume number is **50%** engagement. Reading efficiency **+35%**. Comprehension **2.5×**. Doubts from hours to **3–10 minutes**. **99.5%** uptime. I do not have the study protocol on the resume — not headcount, not a survey instrument. GitHub currently describes **AegisAI** on that repo. I am answering the **PDF product**, not the wrong README.

**Trap.** Invented MAU. Mixing AegisAI. Stealing LangSmith onto GiftedBooks. Claiming 50% was a randomized trial.

### Q. 40% faster reports — faster than what?

**Spoken.** Faster than the same product **before** the custom report builder path I shipped: FastAPI plus NestJS plus PostgreSQL, with the warehouse query not going through a human each time. The resume number is **40%**. I do not have the before/after query log window on the resume. I will not invent p95 or a sample size. **sub-210 ms** is the **dashboard** path, not this 40%.

**Trap.** “Reports went from 10 seconds to 6.” Fake timestamps. Crediting Redis −35% as the report-builder number. Redis is the **bot database** path.

### Q. 65% data-analysis productivity — how do you know?

**Spoken.** Mechanism: people who were not writing SQL could ask the LangChain SQL RAG chatbot and get a result **behind three-tier RLS**, so they stopped waiting on an engineer for every grain of the warehouse. The resume number is **65%**. I do not have the study protocol on the resume. I will not invent tickets closed per week. Isolation is part of why it is usable: a fast leak is not productivity.

**Trap.** “The LLM is 65% more accurate.” Accuracy ≠ productivity. Superuser SQL. LangSmith on this bullet.

### Q. 35% less bot database latency — Redis?

**Spoken.** Yes. Every natural-language turn was hitting Postgres. It felt like “the model is thinking” when the database was the cost. I measured **bot database** latency, put Redis on repeat questions and schema lookups, and did **not** skip RLS on a cache hit. Cached answers are still tier-correct. The resume number is **−35%**. I will not invent a TTL or a p95. Redis **is** evidenced on Ylogx. This is not the Rate Limiter design (independent SDE I live-round count = **1**, not UTA).

**Trap.** Invented TTL. Cache without a tier key. Claiming you shipped a distributed rate limiter at Ylogx. Bigger RDS as the story you actually shipped.

### Q. 99.9% uptime — show me the dashboard.

**Spoken.** That number is on the resume for the Ylogx serving path. I will not invent a monitoring vendor or a twelve-month SLO PDF. The mechanism I **know** and owned is: Dockerized services on **ECS** with a **desired count**, an **ALB** in front with **health checks** so a sick task is not left in rotation, **CloudFront** at the edge for the static/dashboard path, and **CI/CD** so we were not SSHing a snowflake host. Fail a health check → ALB stops sending traffic; ECS replaces the task. That is how I talk about 99.9% without a fake Datadog screenshot. I do not have the incident CSV on the resume.

**Trap.** “I set up Datadog / New Relic / PagerDuty.” Multi-region. Kubernetes as the intern orchestrator. 99.9% on IQVIA or GiftedBooks (GiftedBooks is **99.5%**). Inventing a named outage to sound humble.

### Q. Horizon 40% obstacle detection and 55% collision risk — protocol?

**Spoken.** Mechanism: 60 FPS GStreamer so the operator feed was live; ZED 2 at 2M+ pts/s into occupancy; **costmap fusion** on the sensors we had, not a purchase order for another camera. **+40%** obstacle detection and **−55%** collision risk are resume numbers on that pipeline. I do not have the study protocol on the resume — not a labeled bag count. 17th of 80+ is the **team** result at ERC 2024.

**Trap.** Invented mAP on the rover. Claiming you designed the rocker-bogie. Missed ERC. Argus 24 FPS mixed in.

### Q. Argus 50% fewer violations and 2× compliance — is that mAP?

**Spoken.** No. **73% → 89% mAP** is the detector eval on **15,000+** images. **−50%** violations and **2×** compliance are **site** outcomes after humans got alerts they could trust, logged to Postgres, on **20+** camera feeds at **24 FPS**. I do not have the site study protocol on the resume. I will not invent a plant name or a shift window.

**Trap.** Equating 89% mAP with 89% of pixels or 89% of people. Scaling 73% mAP to 20 cameras. LLM captions as the detector.

---

## 2. Ownership (intern title vs Architected / Built / Designed)

### Q. You wrote Architected. You were an intern. Who actually owned this?

**Spoken.** Intern is the title. I still say **I** for the modules I designed and shipped. At IQVIA I owned the LangGraph wiring, the ranker versus first-hit default, Hybrid plus GraphDB versus vector-only, and LangSmith evals including test-case generation. At Ylogx I owned the report-builder path, SQL RAG behind RLS, Redis on the bot path, and the dashboard deploy path onto CloudFront / ECS / ALB. At Horizon I was **core software**, not team lead: camera, mapping into RViz/Gazebo, costmap. I credit the team for ERC placement and for mechanical / electrical / hardware. I did not hire anyone. I do not have a mentee count.

**Trap.** “I led a team of N.” Fake manager. Taking ERC 17th as a personal trophy without the team. Claiming you designed the rover chassis. Claiming EKS cluster owner.

### Q. What did you *not* own on Horizon?

**Spoken.** I did not own mechanical design, PCB, or the competition logistics. I owned the software path that had to run on the day: GStreamer 60 FPS, ZED 2 occupancy, costmap fusion. 17th of 80+ is **we**. Stale costmap — operator live, planner on last second’s rocks — is the ship-blocker **I** treated as mine.

**Trap.** Hero module. “I built the rover.” Invented missed ERC.

---

## 3. Skills inventory vs evidenced

Say this once if they read the skills block like a menu: *Listed means I have used the tool enough to name it. I will not fake production on a bullet that does not exist.*

| Skill | Spoken (evidenced **or** listed) | Trap |
| --- | --- | --- |
| **Java** | Live Code for this loop is Java. Skills line lists Java. Internships were **Python / TypeScript**; Horizon was **ROS2**; Argus detector is **Python**. Same `HashMap` / heap / graph ideas. I did not ship LangGraph on the JVM. | Spring Boot ran IQVIA. “I am a Java intern.” |
| **TypeScript** | Ylogx **NestJS** (auth, CRUD) plus React dashboards. StratifyLabs is a Next.js product on the resume. | TypeScript wrote LangGraph. |
| **Kubernetes** | **Listed. I will not fake prod.** Ylogx orchestrator I will defend is **ECS + Docker**. | “My EKS cluster.” k8s as the 99.9% mechanism. |
| **GraphQL** | **Listed. I will not fake prod.** Ylogx intern path is **REST**. | GraphQL was the BI API. |
| **ProtoBuf** | **Listed. I will not fake prod.** Not on an intern bullet. | ProtoBuf between FastAPI and NestJS. |
| **Nginx** | **Listed. I will not fake prod.** Ylogx edge I will defend is **CloudFront → ALB → ECS**. | Nginx terminated TLS at Ylogx. |
| **OAuth 2.0 / JWT** | Skills. Ylogx isolation I will defend is **RBAC at NestJS + RLS at Postgres**. JWT as the API identity pattern is honest CS; I will not invent a vendor flowchart. | Payload recipes. Auth0 **as Ylogx prod** (Auth0 is listed; named hop is not “I ran Auth0”). |
| **Auth0** | **Listed. I will not fake prod.** Do not invent Auth0-as-Ylogx-prod. | “We used Auth0 for three-tier RLS.” RLS is Postgres. |
| **Redux** | **Listed. I will not fake prod.** Resume dashboards say **React.js, Recharts**, not Redux. | Redux shipped the 30 KPIs. |
| **Prisma** | **Listed. I will not fake prod.** Warehouse bullet is **PostgreSQL**. | Prisma was the intern ORM for RLS. |
| **Mongo / MySQL / Firebase / Supabase** | **Listed. I will not fake prod** as the intern warehouse. Source of truth I will defend is **PostgreSQL** (Ylogx; Argus event log). | Mongo was Ylogx. Firebase was GiftedBooks (not on PDF). |
| **Redis** | **Evidenced. Ylogx −35%** bot database latency. Cache-aside, tier-scoped keys, RLS still applies. | Redis = CDN. Redis = rate limiter you shipped. Kafka. |
| **TensorFlow** | **Listed.** Argus on the resume is **YOLOv9 + OpenCV**, 73→89 mAP. I will not pretend TF training was the quoted jump. | “I trained YOLOv9 in TensorFlow.” Haar-cascade “we tried first.” |
| **Pandas / NumPy** | **Listed.** Honest as Python data / eval glue around intern and CV work. Not a named production service. | “I built the warehouse in Pandas.” |
| **LangSmith** | **Evidenced. IQVIA.** Traces, evals, test-case generation on Deep Research and Hybrid RAG. | LangSmith on Ylogx / GiftedBooks / Stratify. Logging full confidential BRDs. |
| **FAISS** | **Listed** as Vector DB. IQVIA BRDs are **Azure AI Search Hybrid + Semantic**, not FAISS-prod. | FAISS shipped the BRD system. Qdrant. |
| **Neo4j** | Skills say Graph DB (Neo4j). IQVIA bullet says **GraphDB**. I say GraphDB; Neo4j if they point at skills. Cosine ≠ `DEPENDS_ON`. | “Neo4j cluster I operated.” |
| **OpenAPI** | **Listed** (tools: Postman, OpenAPI Spec). Honest as how I describe REST contracts. Not a billed API-gateway product. | OpenAPI generated the LangGraph. |
| **Docker** | **Evidenced.** Ylogx ECS + Docker; Argus containerized 20+ feeds. | Docker = Kubernetes. |
| **AWS / CloudFront / ECS / ALB / Route 53** | **Evidenced Ylogx.** GoDaddy registrar ≠ Route 53 zone. | I ran EC2 by hand. Multi-region. HTTP/3 in prod. |
| **LangChain vs LangGraph** | Ylogx chatbot = **LangChain SQL RAG**. IQVIA = **LangGraph** (state, fan-out, retry). Do not retrofit LangGraph onto the Ylogx bullet. | “All my RAG is LangGraph.” |
| **WebSockets** | **Listed.** 30 KPIs: describe live-ish REST/poll unless they ask. I will not invent Redis pub/sub fan-out for all 30 tiles. | WebSockets bypassed RLS. |
| **Three.js / WebGL** | StratifyLabs URDF editor with WebGL; GiftedBooks VR labs. | Horizon Gazebo = Three.js. |

### Q. Why is Java on the resume if internships were Python?

**Spoken.** 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/React, ROS2, and YOLOv9 in Python. I will not pretend IQVIA ran on Spring. If this shop wants the same agent graph in Java services, the ideas transfer: nodes, state, retrieval interfaces. Syntax is not the architecture.

**Trap.** “I shipped Java at IQVIA.” Fake Spring Boot.

---

## 4. GiftedBooks GitHub vs PDF; InstaRecon

### Q. I opened github.com/adarshx01/giftedbooks. This README is AegisAI.

**Spoken.** That is a **mismatch**. GiftedBooks on the August 2026 resume is a VR learning suite: 3D labs, AI avatars, RAG on the student’s PDF, **sub-300 ms**, **99.5%** uptime, PYQ topics, doubts **hours → 3–10 min**, reading **+35%**, engagement **+50%**, comprehension **2.5×**. I will answer from the **PDF**. I will not mix AegisAI features into this story. I would fix the README. I will not pretend they are the same product in this room.

**Trap.** Defending the wrong README. Inventing a merge timeline. Stealing IQVIA LangSmith onto GiftedBooks.

### Q. What is InstaRecon / PhiSiFi on GitHub?

**Spoken.** A **security-awareness demo**, with **consent**, not production attacks. I will not walk phishing, credential-capture, or exploit steps. If you want shipped work, StratifyLabs, Argus, Ylogx, or IQVIA.

**Trap.** Payload talk. Turning it into an LP story. “I hacked accounts.”

---

## 5. Why two RAG systems (IQVIA vs Ylogx vs GiftedBooks)

### Q. Why do IQVIA and Ylogx both say RAG? Did you copy-paste?

**Spoken.** Same **shape**: retrieve a small context, then generate. Different **indexes and guardrails**. Ylogx is **SQL RAG**: natural language → schema → SQL → execute **as the user’s database role** so **RLS** applies, Redis **−35%**. IQVIA Deep Research is **open web**: many sites, rank by agreement, refuse without a span. IQVIA Hybrid RAG is **confidential long BRDs**: Azure hybrid plus semantic plus GraphDB hops plus LangSmith. GiftedBooks is a **smaller PDF RAG** for one student’s notes, **sub-300 ms**. I do not retrofit LangGraph onto Ylogx. I do not put LangSmith on Ylogx or GiftedBooks.

**Trap.** “RAG is RAG.” One vector index for all three. Qdrant. Superuser SQL.

### Q. Why two systems at IQVIA? Why not one RAG?

**Spoken.** They fail differently. Deep Research has **no single gold page** — you rank **200+** websites and contradictions are the job. Hybrid RAG has a **gold document** that is **200+** pages — clause IDs, tables, requirement graphs. Stuffing either into a prompt invents citations. Vector-only misses `R-141`. A linear chain cannot retry a dead scrape. So: one graph orchestrator, **two** retrieval backends. Ranker on the web side. Hybrid plus GraphDB on the BRD side. LangSmith on **both**.

**Trap.** “I built two because the JD liked RAG.” Claiming one index served BRDs and the open web.

---

## 6. Line-by-line PDF questions

### Q. What is a BRD?

**Spoken.** A **business requirements document**: long, structured, confidential. Ours were **200+** page BRDs and PDFs. Clause IDs, tables, “this requirement depends on that one.” That is why vector-only was the wrong default and why test-case generation from the **wrong section** is expensive. I do not dump the BRD into the prompt. I do not log the full BRD into LangSmith in production.

**Trap.** “BRD means PDF.” Invented page counts beyond 200+. Claiming you authored the BRD.

### Q. Hybrid vs semantic in Azure AI Search?

**Spoken.** **Hybrid** on Azure is lexical (BM25-like) **plus** vector in **one** query. **Semantic** on that bullet is Azure’s **semantic ranker on top** — not “I invented an embedding.” Lexical hits `REQ-1044` and table names. Semantic hits “what is the retention policy.” Graph hops hit “which requirements depend on this.” I do not say “semantic search” when I mean “we cosined vectors.” FAISS-only misses IDs.

**Trap.** Hybrid = two separate apps. Semantic = GraphDB. Qdrant.

### Q. GraphDB vs vector DB?

**Spoken.** A vector DB retrieves **similar chunks**. A graph stores **relationships** — `DEPENDS_ON`, traces a requirement to children. Cosine similarity is not a foreign key. Skills list FAISS and Neo4j. IQVIA prod retrieval I will defend is **Azure AI Search + GraphDB**. I say GraphDB unless they point at Neo4j on the skills line.

**Trap.** “GraphDB is just embeddings with extra steps.” Neo4j cluster ops you did not do.

### Q. Firecrawl vs Playwright?

**Spoken.** **Firecrawl** is a crawl/scrape **service**: URL in, cleaned markdown/HTML out, good when you want coverage without owning every site’s DOM. **Playwright** is a **browser**: you render JavaScript, click, wait; expensive; necessary when the page is a live app and a static fetch is empty. On Deep Research I run them in **parallel** with Bing and DuckDuckGo because each family fails on different sites. A failed Playwright scrape must **not** wipe Bing hits — that is why the graph is **stateful**. Frugality: fewer Playwright renders, more snippets, when cost spikes.

**Trap.** “Playwright is just a scraper.” Firecrawl on confidential BRDs. Claiming you built Firecrawl.

### Q. ECS task vs ECS service?

**Spoken.** A **task definition** is the recipe: image, CPU, memory, ports, env. A **task** is one running copy of that recipe. A **service** keeps a **desired count** of tasks, attaches the **ALB** target group, and rolling-replaces on deploy. Cluster is the pool. Scale is desired count, not resizing my laptop. Ylogx shipped **ECS + Docker**. I will not invent Fargate versus EC2 launch type, vCPU, or ECR versus Docker Hub.

**Trap.** Task = service. Kubernetes. I ran EC2 by hand.

### Q. ALB vs CloudFront?

**Spoken.** **CloudFront** is the **CDN / edge**: TLS at PoPs, cache static and cacheable GETs close to the user. **ALB** is **L7 load balancing** in the region: host/path rules, target groups, **health checks**, ACM. Ylogx path: client → CloudFront → ALB → ECS (NestJS + FastAPI) → Redis / Postgres. **Do not cache tenant JSON** at the edge. Redis **−35%** is **application** cache next to ECS — different layer. CloudFront 403 ≠ ALB 502 ≠ origin 5xx ≠ RLS 403.

**Trap.** ALB is a CDN. CloudFront does RLS. Nginx as the named edge.

### Q. How can you claim 99.9% without Datadog?

**Spoken.** Resume number, Ylogx serving path. Mechanism I know: **ECS desired count** plus **ALB health checks** plus **CI/CD** plus CloudFront for the static/dashboard path so origin is not the only hop. A failing health check drops the task from rotation; the service starts a replacement. I will describe that. I will **not** draw a fake Datadog dashboard or name a vendor I did not put on the resume. I do not have the SLO window on the PDF.

**Trap.** Invented vendor. 99.9% on IQVIA. Multi-region active-active.

### Q. PPE mAP vs accuracy?

**Spoken.** **Accuracy** on a class label is “of images, how many did I tag right” — the wrong summary for detection. **mAP** is mean **average precision** over PPE classes: for each class, the area under the precision-recall curve after matching boxes by **IoU**. **89% mAP ≠ 89% of pixels** and ≠ 89% of workers. **IoU** = intersection over union versus ground truth. **NMS** collapses duplicate boxes. At **73% mAP** you miss helmets **and** fire junk boxes; either one trains people to ignore alarms. I treated 73% as a fail, trained on **15,000+** images, iterated YOLOv9 to **89%**, and required **24 FPS**. Site **−50%** / **2×** are not mAP.

**Trap.** mAP = accuracy. 89% of people wear PPE. Haar story. LLM detector.

### Q. Why 24 FPS?

**Spoken.** The stream is live industrial video. If infer + NMS + I/O cannot keep **24 FPS**, you either drop frames or you alert on a stale world — same class of bug as Horizon’s stale costmap. **20+** cameras makes a two-stage detector the wrong default. One-stage YOLO is the latency defense. End-to-end budget is decode + infer + NMS + I/O. Prefer drop a stale frame over a delayed “live” PPE alert. Horizon **60 FPS** is **GStreamer video infra** on the rover, **not** this YOLO budget. Stratify **−30%** is ML **iteration time**, not FPS.

**Trap.** Mixing 24 and 60. Claiming 24 FPS is mAP. 20 copies of the fattest model.

### Q. CUSAT 8.42 — is that inflated?

**Spoken.** Cochin University of Science and Technology, B.Tech CSE, October 2022 to May 2026, **CGPA 8.42 / 10**. I will not round it up. It is education, not a Leadership Principle metric. Internships plus ERC plus eight hackathons are the work. If they want Hire and Develop: IEDC CUSAT Tech Team — team member, not a manager, no mentee count.

**Trap.** “Almost 9.” Using CGPA as Earth’s Best Employer. Fake rank in class.

### Q. Why Kerala? Will you relocate to Karnataka?

**Spoken.** I intern in **Kochi** and I studied at **CUSAT**. This role is **ADCI Karnataka**. I will **relocate**. I am not asking for a Kochi exception in this loop.

**Trap.** Drama. Family essay. “Only if remote.” Long weather talk.

### Q. Education dates overlap IQVIA (Apr 2026–present) and May 2026 graduation?

**Spoken.** Final-year CSE with an intern role in Kochi from April 2026. Graduation window on the PDF is May 2026. I will not invent a day-by-day calendar. The intern work is real: two FastAPI systems at IQVIA.

**Trap.** Fake “I already graduated in March.” Hiding the overlap.

### Q. Hackathons — eight events. So what?

**Spoken.** First at **CodeRecet**, Best Project at **MLH.io**, runner-up at **Magnathon 2.0 (IEEE)**, across **eight** national/regional hackathons. Resume wording is rapid software development and production-ready deploy. I will not invent hours-to-prize. This is Bias for Action **backup**, not a fourth Ylogx primary.

**Trap.** Fake prize money. Fake headcount. Using hackathons as Hire and Develop with invented juniors.

### Q. IEDC CUSAT / leadership line?

**Spoken.** Software team member of Team Horizon and Tech Team of IEDC CUSAT. **Not** a people manager. I shared setup and reviews so the next laptop could run the build. No mentee count. No hiring bar I ran.

**Trap.** “I managed the club.” Fake reports.

### Q. StratifyLabs — GitHub is default Next.js.

**Spoken.** Resume wins: **stratifylabs.design**. Computer-vision SaaS: 3D simulation lab, browser inference, **−30%** ML iteration time, marketplace **50+** models/datasets, URDF editor with WebGL, Gemini RAG bots as 3D characters on the **current** sim. I will not invent README extras. Deep RLS stays on Ylogx. No LangSmith on this bullet.

**Trap.** Merging Horizon Gazebo. Invented GTM. 50 GPUs per user.

### Q. Argus GitHub README 404.

**Spoken.** I will not invent routes. Resume: YOLOv9 **73% → 89% mAP**, **15,000+** images, PPE plus attendance, **24 FPS**, **20+** cameras, Postgres logs, containerized, violations **−50%**, compliance **2×**. Purplle-style YOLOv8 + ByteTrack is **prep-only** — not this PDF.

**Trap.** Invented API. Haar first. EKS.

### Q. WebSockets / ProtoBuf / GraphQL are on skills. Where in the bullets?

**Spoken.** Skills inventory. Intern BI path I will defend is **REST**. I will not fake those three as Ylogx production. If they want a real protocol story: Horizon GStreamer is a **pipeline**, UDP-family live video, not GraphQL.

**Trap.** Inventing a GraphQL gateway to sound senior.

### Q. Why Amazon / why SDE I?

**Spoken.** I already own a slice **end to end** — interface, data, isolation, deploy, eval — and I want that ownership where the customer bar is the product. Ylogx: RLS in the database through CloudFront. Horizon: the date did not move. IQVIA: a fluent paragraph is not an answer unless LangSmith can show a span. That is the job I want: one service, mine, correct under load. I will not recite sixteen principle names.

**Trap.** “Amazon is the dream company.” Fake retail HLD.

---

## 7. Metrics table (say the mechanism; do not invent N)

| Resume number | Mechanism you may speak | Do not invent |
| --- | --- | --- |
| 200+ websites | Parallel tools + agreement ranker + refuse without span | nDCG, gold-set size |
| 200+ page BRDs | Hybrid + semantic + GraphDB + citation-gated writer | Token counts, Qdrant |
| 40% faster reports | Custom report builder on FastAPI/NestJS/Postgres | p95, query-log window |
| 99.9% uptime | ECS desired count, ALB health checks, CI/CD, CloudFront | Datadog, SLO months |
| +65% analysis productivity | SQL RAG behind 3-tier RLS | Tickets/week, model accuracy |
| −35% bot DB latency | Redis cache-aside, tier keys, RLS still on | TTL, p95 |
| 60% ops efficiency | 30 live KPI dashboards vs ad-hoc waits | Survey N, hours saved |
| sub-210 ms | Dashboard path, cache + edge + pooling, **not** LLM | Chatbot latency |
| 17th / 80+ | Team ERC result | Personal trophy, missed ERC |
| 60 FPS | GStreamer camera pipeline | Argus 24 FPS mix |
| 2M+ pts/s | ZED 2 → RViz/Gazebo | Shipping all points to a laptop |
| +40% obstacle / −55% collision | Costmap fusion on sensors we had | Rover mAP, extra lidar story |
| 73% → 89% mAP | YOLOv9 on 15,000+ images, eval gate | Accuracy, Haar, 95 mAP |
| 24 FPS | One-stage detector budget × 20+ cameras | 60 FPS rover mix |
| −50% violations / 2× compliance | Alerts after the mAP gate, Postgres audit | Equating to mAP |
| sub-300 ms / 99.5% | GiftedBooks API + hosting | LangSmith, AegisAI |
| +35% reading / +50% engagement / 2.5× | RAG + PYQ + VR labs | Study protocol |
| hours → 3–10 min doubts | Grounded PDF Q&A | Student headcount |
| −30% ML iteration / 50+ marketplace | Stratify browser inference + catalog | GPU farm, GTM |

---

## 8. Overlap, dates, Java, and “too much intern”

### Q. IQVIA April 2026–present and Ylogx through October 2025 — overlap?

**Spoken.** They **do not overlap**. Ylogx ends October 2025. IQVIA starts April 2026. Horizon is February–June 2024. I will not invent a gap-year essay for November 2025–March 2026 unless they ask; I will not fill it with off-resume stories (Django portal, Purplle CCTV, abliteration) unless I have confirmed I want them on the record. Those are **prep-only**.

**Trap.** Leading Dive Deep with SEO 403. Purplle as Argus. Qdrant.

### Q. Three internships in two years. Are you a tourist?

**Spoken.** Each has a different failure mode I can go deep on. Horizon: **stale costmap** versus a live 60 FPS feed, date does not move. Ylogx: chatbot “worked” while **latency and isolation** were the product. IQVIA: **first-hit ranking** and **vector-only BRDs** look like RAG and still ship contradictions. I would rather go deep on one than tour all three in ninety seconds.

**Trap.** Reciting all metrics. Using the same project as primary for a fourth LP.

### Q. You used Architected twice at IQVIA.

**Spoken.** Two systems, one intern. I will not collapse them. Deep Research is web ranking. Hybrid RAG is BRD retrieval plus test-case generation. Shared pieces: FastAPI, LangGraph, LangSmith. Different indexes.

**Trap.** One architecture slide for both, then failing the BRD versus web probe.

---

## 9. Short “if they open the laptop” redirects

- **AegisAI README** → PDF GiftedBooks only.
- **InstaRecon / PhiSiFi** → consent demo, no exploits → Stratify / Argus / Ylogx / IQVIA.
- **argus-stream-api-server 404** → resume metrics, no invented routes.
- **Stratify default Next.js README** → stratifylabs.design / resume.
- **Conv-BI / Warpflow GitHub** → Ylogx-adjacent, not extra resume jobs.
- **Login Tracker** → **unverified**. Do not volunteer as Job **10454435**.
- **Rate Limiter you shipped** → I did **not**. Independent SDE I live count = **1**, OA+3 SD, not UTA. Ylogx Redis is cache.

### Q. RLS vs RBAC — you listed both.

**Spoken.** **RBAC** is which API the caller may hit — NestJS, from the identity. **RLS** is which **rows** Postgres returns even if the SQL is a JOIN or LLM-generated. Resume: RLS and RBAC for **three organizational tiers as role**. NestJS maps caller → DB role; Postgres refuses the row. App-only `WHERE org_id = ?` dies on the first missed JOIN. Scaling ECS **without RLS scales leaks**. Redis keys include the tier. I will not invent the three tier names.

**Trap.** “RLS is a React hide.” Auth0 is RLS. Service-role connection for the bot.

### Q. FastAPI and NestJS in one intern. Why two?

**Spoken.** NestJS owned **auth, RBAC, report CRUD**. FastAPI owned **SQL RAG** (NL → constrained SQL). Postgres is the source of truth. React/Recharts is the 30 KPIs. I will not collapse them into “a Node monolith.” Live Code is still Java.

**Trap.** NestJS wrote LangGraph. FastAPI did JWT as the only security (RLS is the row gate).

### Q. CI/CD / GitHub Actions on skills. What shipped?

**Spoken.** Ylogx: automated CI/CD onto **ECS + Docker**. That is part of how I talk about **99.9%** without a vendor screenshot: we were not patching a snowflake host. I will not invent the YAML, the exact GitHub Actions job names, or a blue/green diagram that is not on the resume.

**Trap.** Fake pipeline file. Kubernetes deploy. “I built GitHub Actions the product.”

### Q. OpenCV vs YOLOv9 vs TensorFlow?

**Spoken.** Argus: OpenCV **decodes and resizes**. **YOLOv9** is the detector. Quoted jump is **73% → 89% mAP** on **15,000+** images, not “I replaced Haar with YOLO.” TensorFlow is **listed**; I will not claim TF training **is** that jump. Stratify is browser inference and a marketplace — different product. Horizon is ROS2 / GStreamer / ZED — not YOLO.

**Trap.** Haar-first story. TensorFlow as the intern detector. Mixing 24 FPS with 60 FPS.

### Q. Local LLMs / OpenAI API on skills?

**Spoken.** Skills. IQVIA and Ylogx bullets name **LangChain / LangGraph / LangSmith / Azure AI Search**, not a foundation model I trained. Stratify names **Gemini** for scene bots. GiftedBooks is PDF Q&A, vendor not named on the PDF. I will not invent which intern called OpenAI versus Azure. Frugality: I did not train a foundation model.

**Trap.** “I fine-tuned GPT.” Abliteration story (prep-only). Qdrant.

### Q. Vercel / Render / Railway?

**Spoken.** Tools line. Ylogx production hop I will defend is **CloudFront / ECS / ALB**. GiftedBooks **99.5%** and **sub-300 ms** are resume hosting numbers; I will not invent the vendor. I will not put Railway on the Ylogx 99.9% sentence.

**Trap.** “Ylogx ran on Vercel.” Mixing 99.9% and 99.5%.

### Q. HTML / CSS / Tailwind / ShadCN / Three.js?

**Spoken.** Front-end skills. Evidenced UI: Ylogx **React + Recharts** (30 KPIs); Stratify **WebGL URDF** / 3D lab; GiftedBooks **VR labs**. I will not fake a design-system ownership story. Three.js/WebGL is Stratify, not Horizon Gazebo.

**Trap.** Redux shipped it (Redux is listed, not in the bullet). ShadCN as a Ylogx metric.

---

*End of resume mock. Numbers only from `Resume_Adarsh_Vishwkarma_Aug26.pdf`. No invented protocols. Job 10454435 still none. Unnamed stays unnamed.*