Technologies & Tools
What I've done
- Building three internal AI systems on Lenskart's AI Engineering team: a hiring intelligence platform, a CEO intelligence brief, and a competitor creative-intelligence platform.
- Made candidate searches survive crashed workers and provider outages. A lease-and-heartbeat state machine over Postgres reclaims a dead worker's in-flight work, and per-provider failure classification falls back across 4 LLM providers mid-search.
- Made candidate ranking explainable and repeatable. A deterministic 4-key comparator (score, evidence coverage, confidence tier, name) replaced one opaque LLM score, and identity dedup sends ambiguous matches to human review instead of auto-merging them.
- Made sure no claim reaches the CEO brief on one weak source. An evidence layer admits a claim only on 1 primary or 3+ independent origins above 0.75 match confidence. Built end to end: 40 source connectors, a 9-tool MCP server backing 18 curation skills, and a measured cost of $1.09 per curation session.
- Replaced a creative-analysis pipeline that fabricated 27% of its rows (78 of 287). In the new 7-layer FastAPI and Next.js platform, a deterministic guardrail checks every generated figure against tool evidence instead of trusting an LLM judge.
- Raised LLM-judge agreement with human labels from 54% to 89% (27 held-out labels in the latest batch) with a tiered judge (deterministic checks, then a classifier, then LLM review), planted-error tests, an overfit audit and a regression gate.




