LLM engineer in India — Deepak Kumar

Swapping in whatever model shipped this month rarely fixes a bad answer; the retrieval step usually does. At India Today Group I built RAG over an editorial archive on MongoDB Atlas Vector Search, where the work that actually moved quality was chunking tuned to how the corpus is written, hybrid retrieval so exact names survive embedding, and a measured evaluation set built from real queries rather than invented ones.

• Chunking and embedding strategy tuned per corpus, then measured — not assumed
• Hybrid keyword + vector retrieval, so proper nouns and rare terms still match
• Evaluation sets built from real user queries before a change is promoted
• Retrieval quality tracked separately from generation quality, so failures are attributable

This site's own blog and its technical books are written by multi-agent pipelines I built and run on a schedule, with no human in the loop per article. That only works because the interesting engineering is in the refusal path: deterministic quality gates that reject a draft before it publishes, a critic pass that is calibrated against its own rubric, and a fail-closed verifier that blocks anything it cannot support from source.

• Deterministic gates — specificity scoring, repetition shingling, length floors — ahead of any LLM judge
• Fail-closed verification: an article nobody could verify does not publish
• Structural fixes over prompt pleading — a model that returns short output needs more slots, not a firmer instruction
• Cost and latency budgeted per call, with resume-from-checkpoint so a failed run is not a full re-spend

An LLM call is a network call to a probabilistic service that bills per token. It needs the things any other dependency needs — a timeout, a fallback, a cost ceiling, monitoring that names the failing component — plus one thing they do not: a way to notice that output quality has drifted when nothing errored. Building that in at the start costs days; retrofitting it after a bad answer reaches a customer costs considerably more.

• Graceful degradation — the product still works when the model is slow, rate-limited or down
• Provider errors distinguished from quality rejections, so a dead API key never reads as a bad draft
• Guardrails on output, plus human review wherever output is published under a real name
• Token and latency cost tracked per feature, alerting like any other service dependency

The retrieval pipeline, the API, the interface and the deployment come from the same person. 9+ years of MERN and Next.js sits behind the AI work, which is usually the difference between an LLM feature that ships and a notebook that gets demoed once.