Backend and data engineer, Bengaluru. I build systems that can tell you when they are wrong: evaluation harnesses, reconciliation gates, measured baselines instead of plausible-looking output.
Portfolio: veer0608.github.io · Résumé: PDF · [email protected]
Open to backend, data, and AI-engineering roles. Bengaluru or relocating.
| agentops | Support agent that takes gated actions behind a policy and escalation gate, with a provider-agnostic LLM seam and a built-in eval harness for tool selection, grounding, cost and latency. |
| citerag | Cite-everything RAG over messy 10-K PDFs with a golden-set eval harness. recall@5 0.37 to 0.767 against a measured 0.85 ceiling. The wins came from fixing PDF text extraction, not clever retrieval. |
| vidsmith | Turns a script into a narrated, captioned video. Uses edge-tts word-boundary timing for exact captions instead of running Whisper after the fact. Deployed and running at vidsmith.duckdns.org. |
| moneytrail | Local-first bank-statement ledger that provably balances. Reconciliation is the first component, not categorisation. If the parse dropped a row, every insight built on it is quietly wrong. |
| reruns | Benchmark for multi-turn support agents, scored on pass^k rather than a single pass or fail. 110 tests. The first measurement run was thrown out on purpose for being unreliable; daily runs since. |
| schemablind | A SQL agent given no schema: four verbs, a database it has never seen, a question. Scored on BIRD execution accuracy, the agent's rows against the reference query's, no judge model. Harness proven (oracle 100%, mute 0%), full model run pending. |
Previously: tested Nostradamus at L&T Finance, an MLOps platform running eight models across EWS, Banking, Self-Cure and Collections. Pipeline validation, SQL verification and UAT on GCP and Kubeflow.
Python · FastAPI · PostgreSQL · pandas · LangGraph · pytest · TypeScript · React