21. LP question bank — which question, which locked story, which talking points

GiftedBooks is a VR learning suite. The customer already has PDFs and still loses hours on doubts. A pretty lab that cannot answer from their material is decoration. I shipped RAG over the upload so retrieve-then-generate ran on their chunks, kept generate small for a sub-300 millisecond API, and added a PYQ topic ranker as analytics separate from the LLM. If there is no supporting span, the assistant abstains. Average doubt time went from hours to three to ten minutes; +35% reading, +50% engagement, 2.5× comprehension, 99.5% uptime. I owned RAG and PYQ, not every 3D lab. Full card: chapter 12 Customer Obsession.

At Ylogx the chatbot returned rows on a demo. Application-only org checks would still let a generated JOIN cross a tier. I was not the named security team; I still owned isolation for three tiers on the shared bot and report path. I put RLS and RBAC in Postgres so SQL ran as the user role, and Redis on repeat questions without skipping RLS. 99.9% held; reports 40% faster; bot DB −35%. Full card: chapter 12 Ownership.

Analysts faced 200+ sites and 200+ page BRDs. The fast invention is a vector index. Similar embeddings are not coverage. I built hybrid + semantic Azure AI Search, GraphDB hops, and LangGraph choosing retrieval mode. Test-case generation is a traced LangSmith job: no overlapping span, no sentence. 200+ sites ranked, 200+ page BRDs with traces. No invented percent. Full card: chapter 12 Invent.

Argus is industrial PPE vision. At 73% mAP the model misses helmets and produces junk boxes. Either one trains people to ignore alerts. I disconfirmed the training-curve belief, iterated YOLOv9 to 89% on 15k+ images, required 24 FPS, then scaled. Violations −50%, compliance 2× after the gate. Full card: chapter 12 Are Right.

I learned LangGraph state and checkpoints so research across 200+ sites could retry a tool without restarting. I learned hybrid retrieval and GraphDB hops on 200+ page BRDs, and LangSmith as the stop condition for generated tests. Tutorial-complete was not architecture. Result is 200+ ranked sites and traced BRDs, not a course badge. Full card: chapter 12 Learn.

I was Tech Team at IEDC CUSAT, a member, not a manager. I did not hire. People did not share one stack; one-head setup made the next event fragile. I paired on the risky module and shared environment plus the failures we had already hit. On Horizon I shared launch order so another software teammate could start perception. I have no mentee count. Full card: chapter 12 Hire.

Argus sits on a safety path. 73% mAP trains alarm fatigue. I required 89% mAP, 24 FPS, NMS, and event logs before more than twenty cameras. Violations −50% and compliance 2× after that bar. The standard was mAP plus FPS plus logs, not a README. Full card: chapter 12 Standards.

The small path is one local GPU loop. I built a browser simulation lab, a marketplace of more than fifty models and datasets, and URDF/WebGL so experiment-and-reuse was the default. Iteration fell 30%. I do not invent users. Backup if they already heard Stratify: ranking research across 200+ sites at IQVIA. Full card: chapter 12 Think Big.

ERC’s date does not move. Waiting for a perfect lab or another sensor would have put useful software after the event. I shipped 60 FPS GStreamer and costmap fusion on the sensors we had, and used field time because Gazebo lighting is not dirt. We placed 17th of 80+. I do not invent a missed ERC. Full card: chapter 12 Bias.

Every natural-language turn hitting Postgres was latency and cost. The obvious move was a larger instance. I measured bot database latency, put Redis on repeat answers and schema metadata, and kept keys scoped to tier so a cached answer could not leak. −35% bot DB latency; 99.9% held. I do not claim a bill or TTL. Full card: chapter 12 Frugality.

I said out loud that a polished lab that cannot answer from the student’s PDF is decoration, and that a generic textbook voice is not a TA. I shipped grounded RAG, abstention, PYQ analytics, sub-300 ms, 99.5% uptime. I also disclose the AegisAI README mismatch and use only the resume product. Hours → 3–10 min. Full card: chapter 12 Earn Trust.

The happy demo said the SQL RAG worked. The path said otherwise: every turn hit Postgres, and a JOIN could cross orgs if isolation lived in the app. I measured bot database latency, put RLS in Postgres, and cached only tier-correct answers. −35% bot DB latency, +65% productivity, 99.9%. I will not lead this with SEO 403. Full card: chapter 12 Dive Deep.

I do not have an interpersonal blow-up on the resume. The conflict I will speak is retrieval. Vector-only is fast; clause IDs and tables need lexical plus graph hops. I argued hybrid + Azure AI Search + GraphDB. When that path was selected, I committed by instrumenting evals instead of reopening the library every week. Full card: chapter 12 Backbone.

From February to June 2024 I was Horizon core software for ERC. The date does not move. I owned GStreamer at 60 FPS, ZED mapping at 2M+ points/s, and costmap fusion. When occupancy lagged a healthy feed, that was the blocker. We placed 17th of more than eighty. I will not invent a miss. Full card: chapter 12 Deliver.

I was a software team member, not a people manager. If only I knew GStreamer topics and launch order, integration stopped when I moved to mapping. I paired until a teammate could launch the node and connect the costmap. Horizon placed 17th as a team. I claim no mentee count. Full card: chapter 12 Earth’s Best.

At more than twenty cameras, wrong alerts scale harm. I did not attach twenty feeds to a 73% model. I moved to 89% mAP at 24 FPS, then scaled, with events in Postgres. Violations −50%, compliance 2× after the gate. Backup if Argus is spent: more ECS tasks without RLS would scale leaks — RLS was on before 99.9% serving. Full card: chapter 12 Success and Scale.