About

At Neatlogs I build the systems that sit on top of agent traces. That means a detection pipeline (a QLoRA-finetuned LLaMA 3.1 model on about 16K labeled samples, 0.83–0.91 F1, served on Azure ML and SageMaker), hybrid BM25 + vector search on ClickHouse, and agents that can walk a trace and draft an eval from context.

In 2025 I was at Thena, where I shipped production agents for a B2B CRM: entity-aware memory, RAG, tool use, and workflows that had to stay idempotent under concurrent ticket load. Chat latency dropped 85%; workflow time dropped 66%. The public product writeups below describe the features I worked on — I did not write those posts.

Earlier: firmware and control-system tooling at Enphase, an LLM-as-judge fellowship at Pi School, and a summer at Medtronic on laparoscopic video masking (US patent application) plus tracking for HUGO RAS. My B.Tech thesis, with Prof. Saurabh Gandhi at CSND Lab, was sleep-stage, lucid-dream, and microsleep classification from EEG.