This chapter is a busted-template drill for Bar Raiser hours. Job 10454435 still none. Unnamed stays unnamed. R1 and R2 already happened. This book is for R3 and R4. Labels: IE-asked Resume-derived.
DO NOT SPEAK AS ADARSH. Every template below is practice-not-yours. If you say it in Zoom as if it happened to you, a Bar Raiser who already has your Aug 2026 resume will catch the contradiction in one follow-up. Genuine answers live in chapter 12 and 05-star-br.md §1–§3. Full fake skeletons (missed SLA, manager fight, mentee of 12, AWS bill, Haar, Qdrant, SEO 403): 05-star-br.md §4. Mechanisms: chapter 14 and chapter 15.
Other SDE I loops asked missing a deadline, last negative feedback, conflict with a manager, Hire and Develop, and “walk me through GitHub.” Those prompts are IE-asked. The failure mode in the room is not “I have no story.” The failure mode is filling the hole with a poster STAR that is not on the resume. Bar Raisers probe metrics, GitHub, and “who said that.” Fake numbers and fake people die there.
Use these eight templates only as a recognition drill: hear the fake, name how it gets busted, switch to the genuine card. Do not rehearse the fake out loud until it feels like memory.
IE-asked LC 8362604 BR missing a deadline; LC 7724048 missed a commitment. The prompt is real. The miss on this resume is not.
Fake template (do not speak). “We missed the European Rover Challenge because the costmap was not ready. I learned to communicate earlier. At Ylogx we also missed 99.9 percent uptime one month and I owned the postmortem.”
How it gets busted. Resume: ERC 2024 17th / 80+. Resume: Ylogx 99.9 percent uptime. There is no miss to cite. The next probe is “what was the actual placement?” or “what was the SLA number you missed?” You then contradict the PDF in the interviewer’s hand.
Genuine switch (chapter 12). I do not have a missed ERC or a missed 99.9 percent, and I will not invent one. What I will speak is how an immovable date was met — 60 FPS and a live costmap sequenced first — and how I would communicate if a ship-blocker were late: early, specific, cut extra hardware, do not cut isolation, do not scale a 73 percent safety model.
Amazon wants risk-flagging, not a fictional failure. “Costmap is planning on stale occupancy; fusion is the path; extra sensor is the cut” is the communication sentence.
IE-asked GFG 2025 BR conflict coworker/manager; GFG April 2026 R3/HM. The prompt is real. A named fight is not on the resume.
Fake template (do not speak). “My manager wanted to ship Friday. I raised my voice in standup. We did not speak for a week. Then I apologized and we became closer. I learned empathy.”
How it gets busted. “What was their name? What exactly did they say? What ticket?” You have no review quote. A Bar Raiser who asks for the other person’s view will watch you invent dialogue. Theatrical humility is not Earn Trust.
Genuine switch. I do not have a named interpersonal blow-up. The conflict I will speak is technical. At IQVIA the fast path is vector-only RAG. Two-hundred-page BRDs have clause IDs cosine misses. I argued hybrid plus GraphDB plus traces. Once we chose the path, I committed on LangSmith evals. Backup: RLS in the database versus app-only. A JOIN leaks. Commit: three-tier RLS as the platform rule. Disagree on retrieval or isolation. Commit on evals or on the platform rule. I will not perform a fight.
“Ship the vector demo by Friday” versus “R-141 must be retrievable” is coverage, not personality.
Haar as a confirmed Argus first attempt is not on the resume. Off-resume unless you later confirm it happened. Do not say it.
Fake template (do not speak). “We started with Haar cascades for helmets. It failed. Then we switched to YOLO and got 89 percent mAP.”
How it gets busted. “Show me the cascade metric. What was Haar mAP? Where is it on the resume?” The PDF says YOLOv9 73 percent to 89 percent on 15,000+ images at 24 FPS. A cascade story adds a first attempt you cannot defend. OpenCV on this product is decode and resize, not the detector you alert on.
Genuine switch. 73 percent mAP on YOLOv9 was the fail. I treated it as a fail, not a blog metric. More than 15,000 images to 89 percent mAP, 24 FPS, then more than twenty cameras with Postgres logs. Violations down 50 percent, compliance doubled, after the gate. Lighting and vest color were dataset work. An LLM describing the frame is not the fix. Fluency is not mAP.
If they ask why not two-stage Detectron: 24 FPS across 20+ feeds. If they ask OpenCV: decode and resize, not the PPE boxes.
Fake template (do not speak). “I built RAG on Qdrant plus a custom embedder. We cut p95 retrieval to 40 milliseconds.”
How it gets busted. Resume IQVIA: Azure AI Search (Hybrid + Semantic), GraphDB, LangGraph, LangSmith. Skills list has Vector DB (FAISS) and Graph DB (Neo4j) as skills, not as IQVIA bullets. There is no IQVIA latency percent. p95 is not on the resume. A follow-up “walk the Azure index types” exposes that you swapped the vendor for a blog stack.
Genuine switch. Retrieval is Azure AI Search hybrid plus semantic plus GraphDB hops. Lexical for clause IDs. Semantic for policy language. Graph for dependencies. Cite or abstain. LangSmith fails a generated test case with no overlapping span. Result: 200+ page BRDs, 200+ sites ranked. No extra percentage. I did not use Qdrant.
FAISS on the skills list is not an excuse to rename Azure. If they ask FAISS: GiftedBooks-shaped local vectors are a different product; IQVIA bullet is Azure. Do not steal GiftedBooks into IQVIA.
IE-asked Bhavya helped juniors; LC 6475219 helped a peer. Honest version is IEDC / Horizon sharing. No mentee count on the resume.
Fake template (do not speak). “I mentored twelve juniors at IEDC. Three got internships. I ran a hiring bar. I improved their ratings.”
How it gets busted. “How did you measure the twelve? What was the rating system? Were you a manager?” Resume: Tech Team of IEDC CUSAT; Horizon software team member; eight hackathons. No reports. No Career Choice. No Amazon-style hiring bar. Invented headcount is the fastest Hire-and-Develop fail.
Genuine switch. I was on the Tech Team at IEDC CUSAT. I did not hire anyone. I do not have a mentee count. I shared environment setup, reviews, and the failure we already hit. At eight hackathons I paired on the risky module. On Horizon I shared GStreamer launch order and costmap constraints so the rover was not one-head. 17th of more than eighty is a team score. CGPA 8.42 is education, not this LP’s metric.
Earth’s Best Employer is the same honesty with a different bar: can they launch perception without my laptop. Still no headcount.
Fake template (do not speak). “Redis TTL was five minutes. p95 dropped from 800 ms to 120 ms. We saved two thousand dollars a month on RDS. I picked db.r6g.xlarge.”
How it gets busted. Resume numbers: −35 percent bot database latency, sub-210 ms dashboards, 99.9 percent uptime, 40 percent faster reports. No TTL. No p95. No bill. No instance class. “What was the CloudWatch dashboard?” forces a screenshot you do not have. Frugality on this resume is Redis versus a bigger RDS as the first move, plus no LLM on the dashboard path — not a finance story.
Genuine switch. Uncached SQL RAG hit Postgres every natural-language turn. I cached schema metadata and repeat answers. A cached answer still had to be tier-correct. The claim I will defend is minus 35 percent bot database latency. Dashboards stayed sub-210 milliseconds without an LLM on that path. I will not invent a TTL, a p95, or an AWS bill.
Why Redis not bigger RDS? Repeat NL and schema lookups are cache-shaped. Why still RLS after cache? A cached leak is still a leak. Cache key includes tenant, not raw NL.
GiftedBooks GitHub currently describes AegisAI. Resume wins. Mixing the README into STAR is how you fail Earn Trust on the artifact they just opened.
Fake template (do not speak). “GiftedBooks is AegisAI, an agent platform with tool calling and a multi-tenant control plane. We had N students on campus.”
How it gets busted. They open github.com/adarshx01 and read AegisAI copy on the giftedbooks repo. Then they ask for the VR labs, PYQ, sub-300 ms, 99.5 percent, hours to 3–10 minutes — the resume product. You either abandon the README or abandon the PDF. Student headcount is also not on the resume.
Genuine switch. GiftedBooks on the resume is a VR learning suite: 3D labs, AI avatars, RAG on the student’s PDF, PYQ topic suggestions. Doubts hours to 3–10 minutes. Sub-300 milliseconds, 99.5 percent uptime, plus 35 percent reading, plus 50 percent engagement, 2.5 times comprehension. GitHub currently mismatches AegisAI on that repo. I will not mix that product into this story. The resume is the source. I will not invent student headcount.
If they have the tab open: one sentence — README currently describes a different product; I will speak the Aug 2026 resume — then continue the student story.
InstaRecon / PhiSiFi is not an LP story. Ethics one-liner, then redirect. No phishing, credential, or exploit steps.
Fake template (do not speak). “For Customer Obsession I OSINT’d Instagram at scale. I captured sessions. I showed how to phish. That is Dive Deep.”
How it gets busted. Amazon will not reward a credential-capture story. It also is not on the intern bullets. A security-aware interviewer will ask for consent, scope, and production impact. You either describe attacks you should not describe, or you look like you used a hobby repo as a fake LP.
Genuine switch. If they open that GitHub: it is a security-awareness demo, consent, no production attacks. Then I move to StratifyLabs, Argus, Ylogx, or IQVIA. Customer Obsession I will speak is GiftedBooks grounded PDF Q&A. Dive Deep I will speak is Ylogx Redis plus RLS. I will not walk exploit steps.
Redirect in one breath. Do not “just a little architecture” of InstaRecon. Chapter 17 if they critique GitHub mismatches.
| Fake (never speak) | Bust probe | Genuine chapter 12 card |
|---|---|---|
| Missed ERC / missed 99.9% | Placement? SLA number? | Deadline without a miss; sequence + flag risk |
| Manager shouting match | Name, quote, ticket | Hybrid vs vector, or RLS vs app-only |
| Haar first, then YOLO | Haar mAP on resume? | 73% YOLOv9 was the fail |
| Qdrant + p95 | Azure hybrid on the PDF | Azure AI Search + GraphDB + LangSmith |
| Twelve mentees | Rating system? Reports? | IEDC / launch files; no headcount |
| TTL, bill, instance class | CloudWatch? Invoice? | −35% bot DB latency; sub-210 ms |
| AegisAI as GiftedBooks | Open GitHub | Resume VR + PDF RAG metrics |
| InstaRecon LP | Consent / production | Ethics line, redirect |
Ylogx SEO 403 as Dive Deep. Prep-only. The anecdote-versus-metric story is bot latency plus JOIN leak. If you lead with www versus non-www, they will think you do not have a production bug.
Kafka / EKS / shard as if you ran them. Skills list has Kubernetes. Ylogx shipped Docker + ECS + CloudFront. I did not shard. I did not operate Kafka. Pretending JVM on internships is the same class of bust: Live Code is Java; production was Python / TypeScript / ROS2.