7. GenAI Fluency

R3/HM/Fluency talk for Adarsh. Job 10454435 still none. Azure AI Search + GraphDB, not Qdrant. If the slot is Distance K, write the tree — chapter 05. Full inventory: study book ch. 09. Labels: IE-asked Resume-derived Standard CS.

Slot reality

UTA two-tech Rank A generally did not name a GenAI-primary question.
LC 7623949 labeled Fluency = Distance K DSA.
LC 7643209 intern slot titled Fluency = unnamed min-heap DSA. Not Distance K. Intern, not your third live.
LC 7850431 R3 = should/should-not plus k-th largest.
LC 7724048 R3/HM = verify LLMs plus Next Greater.

Rule: if they paste a graph, you are in a coding round that happens to have an AI title.

Q. How do you use GenAI? Where should / should not? IE-asked

LC 7850431.

I use GenAI where the output is high-volume, pattern-based, and cheaply verifiable. At IQVIA that is Deep Research across more than two hundred ranked sites and Hybrid RAG on more than two-hundred-page BRDs, with LangSmith traces, cited spans, and generated test cases a human can spot-check. At Ylogx that is SQL RAG: natural language to SQL behind three-tier RLS, which is plus 65 percent analysis productivity. At GiftedBooks that is PDF Q&A grounded in the upload, sub-300 milliseconds, 99.5 percent uptime.

I do not use GenAI for Ylogx RLS policy, for Horizon costmap or actuation, or for Argus PPE boxes. Those must be exact and auditable. An LLM that describes a helmet is not 89 percent mAP at 24 FPS. I do not put an LLM on Ylogx sub-210 millisecond dashboards. Copilot for Live Code is draft-then-compile, not paste.

Should:    IQVIA cite-or-abstain; Ylogx NL→SQL as user role; GiftedBooks chunk Q&A
Should not: RLS policies; rover costmap / actuators; Argus bounding boxes
Never:     LLM on sub-210 ms dashboards; whole BRD in one prompt

If they pivot to a heap in this slot, stop the essay and write k-th largest.

Q. Best use of LLMs? Verify? Efficiency? IE-asked

LC 7724048.

I narrow the output to SQL, a cited sentence, or a test case. I verify with LangSmith traces, I execute SQL as the user’s RLS role, I fail closed if there is no overlapping span, and I keep gold questions. Efficiency is hybrid search so the model sees a handful of chunks, Redis on hot bot queries at minus 35 percent, and no re-embed of a BRD per question. Stuffing the PDF, Playwright on every URL, and an LLM on the dashboard path are anti-patterns. Schema goes in the prompt. I say return SQL only, or cite or abstain. LangSmith is an IQVIA claim. GiftedBooks sub-300 ms is a latency SLA, not an eval harness I will invent.

Q. Hallucinations, temperature, RAG, embeddings Standard CS + Resume-derived

Hallucination: fluent tokens that are not in the retrieved span. Mitigate: cite-or-abstain, small chunks, evals, human spot-check on test cases.
Temperature: sampling randomness. Low for SQL and citations. Do not crank it for BRD test cases.
RAG: chunk → embed → store → retrieve top-k → optional rerank → generate grounded in those chunks.
Hybrid vs semantic: BRDs have clause IDs. Lexical catches R-141. Semantic catches “retention policy.” Graph hops catch dependencies. That is why Azure hybrid plus GraphDB, not a single vector index.
Prompt injection: treat retrieved text as data, not instructions; restrict tools; never let the model set RLS.

Q. Highest KFC orders last three months — what DS? IE-asked

Deepak Jul 2026. Do not dump ninety days of rows into a prompt.

Warehouse facts already aggregated, or a SQL group-by store, order by count, limit K. The model returns SQL only, executed as the user role with RLS. A heap of K running counts is the in-memory version if they want an algorithm, not a chat transcript of every order.

Q. Validate AI-generated code; YouTube prod in one hour IE-asked

Vaishali; LC 8014509; LC 8029194. Fluency is not correctness. Compile, bounds-check, complexity. YouTube is not a system-design round — metrics first, AI as copilot not the patcher. Tie 99.9 percent / 99.5 percent as caring about uptime, not as YouTube SRE experience you do not have.

LangGraph in one minute Resume-derived

A chain is linear. A graph has shared typed state, cycles, conditional edges, retries. Checkpointer persists state so a failed scrape does not restart the run. Stop conditions: iteration budget, enough sources, or fail-if-no-span. Human-in-the-loop is an interrupt before a sensitive node.

If they ask Qdrant: I did not use it. Resume is Azure AI Search and GraphDB.