ML systems engineer in New York. I build speech and language systems that run in production — and study where they quietly go wrong.

MS Computer Engineering (ML Systems) at NYU Tandon. Co-founder and lead ML engineer at Zotivo AI, where I work on production voice AI.

Selective faithfulness in LLM hiring decisions — when a model is told to favour a demographic group, does its written justification say so? A four-arm crossed design over three model families, with blind re-scoring and a two-tier verbalization detector validated against hand labels. Under revision for FAccT 2027; the collection and analysis pipeline is public. → selective-faithfulness

Cultural bias in vision–language models — a pilot audit of LLaVA-NeXT on South Asian representation. Found a 9× domestic-association gap mediated by clothing, and a 14× exoticisation gap on cultural-event imagery. Proposed MS thesis direction. → vlm-cultural-bias-pilot

Reviewer for AIES.

Music recommendation, end to end — the data and feedback-loop half of a four-person MLOps build on a self-hosted music server: a Go scrobbler emitting real listening events, session datasets built for GRU4Rec/SessionKNN, and drift monitoring on Prometheus + Grafana closing the loop back into retraining. → navidrome-mlops-data-proj05

Python Go PyTorch ONNX Runtime FastAPI Kubernetes Docker Prometheus Grafana Redis Parquet PostgreSQL

Speech and voice AI · recommender systems · LLM evaluation and auditing · model serving and drift monitoring

_{Most of my day-to-day production work is closed-source. The repos above are the ones I can show: independent research and coursework, documented so the design decisions are readable without me in the room.}