DO NOT SPEAK AS ADARSH. This chapter teaches Adarsh
Vishwakarma to recognize eight fictional STAR shapes and predict how a Bar
Raiser would break them. It is not material to memorize or perform. Every
fictional statement is isolated inside a .dont block; every
truthful redirect points to a complete genuine card in chapter 12.
Job 10454435, SDE I AUTA APJ, ADCI Karnataka: no public experience still names a live-round question for this Job ID. R1 on 13 Aug 2026 and R2 on 18 Aug 2026 are done. OA is not a round. This notebook prepares only for a possible R3 or R4 without forecasting whether either occurs or what format it uses.
Say the language distinction once: Amazon Live Code is Java. Production internships and projects used Python with FastAPI, LangGraph, YOLOv9 and ROS2, TypeScript with NestJS and React, and ROS2 C++/Python on Horizon. Ylogx and IQVIA did not run on the JVM.
Read the reported prompt, notice the tempting fictional shape, and stop before rehearsing it. Ask which artifact a Bar Raiser would request: an availability log, a manager’s exact words, a rating system, a cloud invoice, event records, an evaluation run, Search Console evidence, or a retrieval trace. Then state the contradiction and move to the linked genuine card.
Labels are evidence boundaries: IE-asked means another candidate reported the prompt family, not that Job 10454435 will ask it. Resume-derived marks the truthful redirect. Standard CS alternative marks general reasoning, never a claim that Adarsh shipped that alternative.
The closing sentence for every fictional shape is: “That event is not on my resume, so I will not present it as mine. The genuine example I can defend is…” Then use the linked chapter 12 card and its locked metrics.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. Do not turn its wording into autobiographical memory.
The reported family is “Tell me about a time you failed to deliver the expected result,” “missed a commitment,” or “missed a deadline.” The fictional mapping below tries to manufacture an operations failure. The truthful mapping is Argus’s 73 percent mAP as a failed quality bar, not a missed Ylogx uptime target.
At an unnamed company, the on-call dashboard promised 99.95 percent monthly availability.
Restore the SLA after a region blip.
I rolled a configuration, wrote a postmortem, and added a monitor.
We missed the month, and I said that I “learned communication.”
The interviewer asks for the availability calculation, the affected region, the monitor definition, the configuration change, the postmortem owner, and the monthly error-budget record. None exists for Adarsh. A claim of 99.95 percent then conflicts with the resume’s actual Ylogx result: 99.9 percent uptime held. The story collapses before the lesson because its primary artifact and metric are fictional.
The Aug 2026 resume does not contain a missed SLA, a region blip, a postmortem, or a missed Ylogx month. It reports 99.9 percent uptime as a result. It also does not report a missed ERC. The source explicitly says to use Argus’s 73 percent mAP as the failed bar when asked for failure.
Redirect to Are Right, A Lot — Argus mAP forced a changed belief. That complete card has the defensible failure, personal action, measurement, discarded option, credit boundary, and 10× answer.
Concrete genuine example. A missed helmet is a silent failure and a junk box is a loud failure. At 73 percent mAP, both made the detector unfit for a human alert. Adarsh used 15,000+ images, moved YOLOv9 to 89 percent mAP, required 24 FPS, and scaled only after the gate.
Q1. What failed if no SLA was missed? The Argus quality bar failed at 73 percent mAP; that is an evaluation failure, not a fabricated calendar or availability miss.
Q2. Who gave the feedback? The evaluation metric supplied the feedback; there is no manager quote to invent.
Q3. What did you personally do? I owned training, mAP and FPS gating, NMS, and event logging; I do not claim an unnamed on-call role.
Q4. How do you know the result? The resume numbers are 73 to 89 percent mAP, 15,000+ images, 24 FPS, 20+ cameras, 50 percent fewer violations, and 2× compliance.
Q5. What would you do differently? I would hold out a camera earlier and label that as a would-add test, not as something already shipped.
Q6. What happens at 10×? I would preserve the quality gate before adding stream workers or batching; I would not multiply a 73 percent detector.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. No dialogue, person, or escalation is real.
Other candidates reported conflict-with-manager, disagreement, and convince- someone prompts. The fictional mapping turns Backbone into a confrontation. The truthful mapping is IQVIA hybrid retrieval versus vector-only retrieval, followed by commitment to the selected eval path.
My manager wanted to ship on Friday, while I wanted tests.
Stand my ground.
I raised my voice in stand-up, escalated to a skip-level, and “won.”
We shipped the week after, and the manager respected me.
The interviewer asks the manager’s name, the exact objection, what the manager would say, which tests blocked Friday, who made the final decision, and what “respected” changed. Adarsh has no transcript or named interpersonal blow-up. Any answer requires inventing more dialogue, and the supposed learning becomes a performance rather than evidence of respectful challenge and whole commitment.
Neither the resume nor the source supports a raised voice, a skip-level escalation, a delayed Friday release, or a manager’s changed opinion. The source locks the conflict to hybrid versus vector-only or database RLS versus app-only filtering, with no fight scene and no quote.
Redirect to Have Backbone; Disagree and Commit — IQVIA hybrid versus vector-only. The full card explains the coverage disagreement, the selected design, and how commitment became visible through LangSmith evals and test generation.
Concrete genuine example. A semantic search can answer “retention policy” yet miss the exact identifier “R-141.” Lexical retrieval handles the ID, semantic retrieval handles policy language, and a GraphDB hop follows dependencies. A generated test without an overlapping span fails.
Q1. What was the disagreement? It was coverage versus fast demo scope: vector-only was simpler, while long BRDs also needed lexical matches and dependency hops.
Q2. What did the other person say? I do not have a transcript, so I state the trade-off without inventing their words or motives.
Q3. What did you personally do? I articulated the retrieval failure mode and instrumented the chosen hybrid path with traces and test-case checks.
Q4. How did you commit after disagreement? I stopped relitigating tools and committed by tracing and evaluating the selected Azure AI Search, GraphDB, and LangGraph path.
Q5. How do you know it worked? The resume supports 200+ page BRDs, evals, and generated test cases; it does not support an invented win percentage.
Q6. What would you do differently? I would place the eval table in the discussion earlier and freeze the evidence slice before debating another tool.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. Its pod, ratings, promotions, and retention result do not belong to Adarsh.
Other candidates reported helping-a-junior, helping-a-peer, and Hire and Develop prompts. The fictional mapping inflates student-team participation into formal management. The truthful mapping is capability sharing through IEDC, hackathon work, and Horizon launch knowledge, with no mentee count.
I led a pod of twelve juniors.
Hire and develop the pod.
I ran one-to-ones, wrote ratings, and promoted two people.
Retention increased 20 percent.
The interviewer asks for reporting lines, the rating framework, promotion authority, one person’s before-and-after behavior, and the retention denominator. Adarsh was not a manager and has no such records. The numerical headcount invites twelve separate identity and outcome probes, while the twenty-percent result requires a population and period that do not exist in the resume.
The resume says IEDC CUSAT Tech Team, eight hackathons, and Horizon core software. It does not say manager, hiring owner, pod lead, twelve mentees, ratings, promotions, or retention improvement. The source explicitly locks the honest answer to sharing setup, review, risky-module work, and launch knowledge.
Redirect to Hire and Develop the Best — IEDC CUSAT Tech Team. That card keeps credit at team-member scope and explains what independent capability looked like without manufacturing a people metric.
Concrete genuine example. If only Adarsh knew the GStreamer topic names and launch order, rover integration stopped when he moved to mapping. Pairing until another teammate could launch the node transferred a concrete capability; it did not create a fictional promotion.
Q1. How many people did you manage? I managed no direct reports and do not have a mentee count.
Q2. What development mechanism did you use? I shared environment setup, reviewed work, paired on risky modules, and transferred Horizon launch and costmap knowledge.
Q3. How did you know another person was unblocked? The concrete check was that another teammate could launch the relevant node and integrate the path without depending on my laptop.
Q4. What outcomes can you defend? I can defend eight hackathons, CodeRecet first place, MLH Best Project, Magnathon 2.0 runner-up, and Horizon’s team finish of 17th out of 80+.
Q5. What was your individual contribution? I shared setup, reviews, and integration knowledge; team awards and ERC placement remain “we” outcomes.
Q6. What would you do differently? I would write a short runbook earlier while being clear that the resume does not claim an existing formal wiki or program.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. The amount and finance response are invented.
The reported family asks about Frugality, reducing cost, or doing more with less. The fictional mapping substitutes a large dollar result for mechanism. The truthful mapping is Ylogx Redis on a cache-shaped database path, with CloudFront and ECS as the serving path and no claimed invoice.
Cloud spend was out of control.
Frugality.
I rightsized instances and saved forty thousand dollars.
Finance sent a thank-you.
The interviewer asks whether forty thousand dollars is monthly or annual, which services changed, the before-and-after utilization, who owned the account, where the invoice sits, and who in finance verified attribution. Adarsh cannot produce any of those facts. “Rightsized” then exposes the absence of an instance class, workload curve, or billing baseline.
The resume contains no AWS bill, dollar saving, finance message, or corporate cloud-ownership claim. It supports Redis reducing bot database latency 35 percent, sub-210 millisecond dashboards, 99.9 percent uptime, and CloudFront/ECS/Docker/ALB/Route 53. The source says never convert those facts into a dollar story.
Redirect to Frugality — Ylogx Redis before a larger database. That full card explains the cache-shaped workload, retained RLS boundary, measured result, and honest limits.
Concrete genuine example. A repeated natural-language question and stable schema metadata can hit a tier-scoped cache. A miss still runs through FastAPI SQL RAG to Postgres as the user role, and the sub-210 millisecond dashboard path does not call an LLM.
Q1. How many dollars did you save? I do not remember or claim a bill; the resume number is a 35 percent reduction in bot database latency.
Q2. Why Redis instead of a larger database? Repeat questions and schema metadata are cache-shaped, so measuring and caching that hot path was the first move.
Q3. What did you personally ship? I shipped Redis on the bot path and preserved tier-correct authorization; I did not own corporate finance.
Q4. Why keep RLS after caching? A cached answer for the wrong organization is still a leak, so cache scope cannot replace database policy.
Q5. What would you do differently? I would document invalidation rules with RLS and add denied cross-tier cache cases from day one.
Q6. What happens at 10×? I would increase safe cache hits, use connection pooling, and keep static delivery efficient before considering a larger database; I did not shard.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. Horizon did not miss ERC in the resume.
The reported prompt family includes strict deadlines, missed commitments, pressure, and what could have been done better. The fictional mapping adds a dramatic event miss. The truthful mapping is delivery on a fixed date despite stale occupancy and field-versus-simulation risk.
On challenge day, the rover was not ready.
Complete a hero weekend.
I pulled three all-nighters, and we barely made it.
We missed a run but “learned.”
The interviewer asks which ERC run was missed, the schedule, what “barely” means, why three all-nighters were necessary, who approved the risk, and how a missed run produced the recorded placement. The resume instead says 17th globally out of more than eighty teams. The invented miss creates a direct artifact conflict with the competition result.
The resume does not say Horizon missed ERC, nearly failed to appear, missed a run, or relied on all-nighters. It says GStreamer at 60 FPS, ZED 2 at 2M+ points per second, 40 percent better obstacle detection, 55 percent lower collision risk, and 17th out of 80+. The source keeps stale costmap as a setback that was handled, not an event that was missed.
Redirect to Deliver Results — Horizon 17th of 80+. The complete card distinguishes activity from the live feed and costmap the judges could actually score.
Concrete genuine example. The operator can see a healthy 60 FPS feed while the planner uses last second’s rocks. Adarsh treated stale occupancy as the ship-blocker, downsampled visualization, kept useful density local, and placed fusion before the actuators.
Q1. Did you miss ERC or a competition run? No such miss appears on the resume, so I will not claim one.
Q2. What was the real setback? The real failure class was a live-looking operator feed paired with stale occupancy, plus field lighting that did not match Gazebo.
Q3. What did you personally own? I owned GStreamer, ZED mapping into the costmap, and fusion on the software path; the 17th-place finish belongs to the team.
Q4. How do you know the result? The resume supports 17th/80+, 60 FPS, 2M+ points per second, 40 percent better obstacle detection, and 55 percent lower collision risk.
Q5. What would you do differently? I would name the ship-blocker in week one, calendar it earlier, and time-box extra hardware versus fusion against collision risk.
Q6. How would you flag a late risk? I would say early and specifically that the costmap is stale, fusion is the path, and extra hardware or polish is the cut.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. Haar, YOLOv8, and 95 mAP are not Argus facts.
The reported family includes “Are Right,” critical feedback, a shortcoming, or a difficult implementation. The fictional mapping supplies a simplistic before-and-after technology swap. The truthful mapping is a measured YOLOv9 quality and throughput gate.
We used Haar cascades for PPE, and they failed before I “switched to deep learning.”
Are Right.
I replaced a one-line OpenCV approach with YOLOv8.
The model reached 95 mAP.
The interviewer asks for the Haar features, cascade training set, baseline metric, false-positive behavior, reason for YOLOv8, and evaluation that yielded 95. None can be reconciled with the resume. The PDF names YOLOv9, 73 to 89 percent mAP, 15,000+ images, and 24 FPS. One model-version probe exposes the substitution immediately.
Haar cascades are absent. YOLOv8 is not the Argus detector. Ninety-five mAP is absent. The supported path is YOLOv9 from 73 to 89 percent mAP, NMS, more than 15,000 images, 24 FPS, Postgres event logs, and scale to 20+ cameras after the gate. OpenCV may decode and resize; it is not a claimed cascade origin.
Redirect to Are Right, A Lot — Argus mAP forced a changed belief. That genuine card starts with the actual mistaken bar: treating 73 percent as progress suitable for production.
Concrete genuine example. NMS prevents duplicate boxes from becoming duplicate alarms, while the 24 FPS gate ensures correct boxes arrive in time. Detection events go to Postgres; full video does not.
Q1. Did Argus use Haar cascades? Haar is not on the resume or confirmed source, so I do not claim it.
Q2. Which model did you ship? The resume names YOLOv9, not YOLOv8.
Q3. What was the actual quality progression? The supported progression is 73 to 89 percent mAP on more than 15,000 images, with a 24 FPS serving gate.
Q4. What did you personally do? I owned training, evaluation, NMS, FPS gating, and event logging.
Q5. How did you avoid scaling the wrong result? I did not move to 20+ cameras until the 89 percent mAP and 24 FPS bars were met.
Q6. What would you improve next? I would hold out a camera earlier and publish concrete missed-helmet and false-box cases, clearly as future evaluation improvements.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. It is prep-only and not the Ylogx deep dive.
Other candidates reported deep-diving a bug, investigating a discrepancy, and finding a problem. The fictional mapping uses a search-indexing anecdote. The truthful mapping is Ylogx’s chatbot demo disagreeing with database latency and cross-tier isolation.
The www and non-www variants returned 403, and Search Console dropped.
Dive Deep.
I found noindex, fixed DNS, and restored ranking.
Traffic increased by “N percent.”
The interviewer asks for the affected hostname, HTTP evidence, Search Console graph, DNS record, indexing date, and the value and baseline behind “N.” The source deliberately leaves the metric as a placeholder. The story has no resume artifact and diverts from the far deeper production mechanism that Adarsh can defend at Ylogx.
The resume does not provide an SEO incident, 403 recovery, noindex change, or traffic metric. The source calls it prep-only and locks Dive Deep to Redis plus RLS: every natural-language turn reached Postgres, while an app-only check could be outrun by a generated JOIN.
Redirect to Dive Deep — Ylogx anecdote versus latency and isolation. The full card contains the measurement, database role boundary, cache mechanism, honest result, and next tests.
Concrete genuine example. A generated JOIN can be valid SQL and still expose another organizational tier. Running it as the user’s database role lets Postgres RLS refuse unauthorized rows; only the allowed result may enter a tier-scoped Redis key.
Q1. What anecdote disagreed with the metric? The chatbot returned SQL and rows in a demo, while database timing showed every natural-language turn still reached Postgres.
Q2. What security detail did you inspect? I inspected where generated SQL ran and moved three-tier isolation into Postgres RLS and RBAC.
Q3. Why not rewrite the prompt first? The measured failure was database work and authorization placement, not proof that generation was the bottleneck.
Q4. How do you know the change mattered? The resume reports 35 percent lower bot database latency, 65 percent higher analysis productivity, 40 percent faster reports, and 99.9 percent uptime.
Q5. What would you do differently? I would add denied cross-tier query and cache-key fixtures from day one.
Q6. What happens at 10×? I would increase safe cache hits and pooling while retaining RLS; ten times the traffic without policy is ten times the leak.
DO NOT SPEAK AS ADARSH. The complete STAR below is fictional practice from §4 of the source. Every person, vendor, size, latency, TTL, and title inside the block is deliberately fake.
This catch-all shape tries to answer complex-project, conflict, evaluation, latency, and ownership prompts at once. The truthful mapping is IQVIA Hybrid RAG with Azure AI Search, a GraphDB, LangGraph, LangSmith, 200+ ranked sites, and 200+ page BRDs—with no invented percentage or gold-set size.
My intern rival hid the Qdrant collection; the gold set had ten thousand labeled BRD questions, p95 was 47 milliseconds, and cache TTL was 90 seconds.
Everything.
I confronted the intern rival.
I became team lead.
The interviewer asks for the collection schema, Azure-versus-Qdrant decision, the ten-thousand-question labeling process, p95 harness, TTL rationale, access history, the other intern’s view, and the title change. Every follow-up creates another contradiction. The resume names Azure AI Search Hybrid plus Semantic and GraphDB, offers no latency SLA or gold-set size, and does not name a team-lead promotion.
There was no Qdrant collection in the resume claim, no rival, no hidden data, no 10,000-question gold set, no 47 millisecond p95, no 90-second TTL, no confrontation, and no team-lead title. The source explicitly lists each as a fact never to merge into the genuine IQVIA story.
Redirect to Invent and Simplify — IQVIA Hybrid RAG and test-case generation. That complete card explains parallel research tools, checkpointed graph state, adaptive retrieval, citation gating, and honest outcomes.
Concrete genuine example. “What depends on R-141?” first needs lexical retrieval for the exact identifier, then a GraphDB hop for dependent clauses. If the generated test case has no overlapping source span, the traced path fails instead of inventing coverage.
Q1. Which retrieval system did IQVIA use? The resume names Azure AI Search Hybrid plus Semantic with a GraphDB and LangGraph adaptive retrieval.
Q2. How large was the gold set? I do not claim a gold-set size; the resume supports LangSmith evals and generated test cases, not ten thousand labeled questions.
Q3. What latency did retrieval achieve? The resume gives no IQVIA p95 or latency SLA, so I do not invent one.
Q4. What did you personally build? I owned the LangGraph nodes, retrieval-mode choice, ranking and traced test-generation path; I did not build Azure AI Search itself.
Q5. How do you know the path worked? The supported outcomes are 200+ ranked sites, 200+ page BRDs, traces, evals, and test-case generation, with no extra percentage.
Q6. What would you do differently? I would freeze an eval slice with expected citations before adding a fourth search tool and make the citation requirement structural earlier.
Q7. What changes at 10×? I would cache scrapes, use fewer full Playwright renders, retrieve small evidence sets, and scale evaluation before scaling fluent output.
Do not invent a missed SLA, manager fight, mentee count, AWS bill, missed ERC, Haar origin, SEO recovery, Qdrant stack, gold-set size, p95, cache TTL, student headcount, or team-lead title. Do not add Kafka, EKS, or sharding as shipped. Do not convert a skills-list item into a project fact.
GiftedBooks uses the Aug 2026 resume product despite the current AegisAI README mismatch. InstaRecon receives one line only: security-awareness demo, consent, no production attacks; then redirect. Login Tracker remains unverified. Rate Limiter has one independent SDE I live count, not UTA.