Senior Bioinformatics Engineer at NYU Langone. 10 years of multi-omics research; now building agentic AI systems for genomics — architectures where the filesystem is the state machine, stage contracts constrain the agent, and scientific decisions stay with the human.

🔭 Flagship: GARS — Genomics Agentic Research System — reproducible bioinformatics workflows through an LLM agent on HPC · ▶ try the interactive demo — recorded real runs, playable human gate, in the browser.

🔌 Tools: HiC-MCP — Hi-C / 3D-chromatin analysis for AI agents. An MCP server exposing the open2c stack (cooler, cooltools) over local contact matrices, with a real Micro-C dataset bundled so it runs offline.

🔬 Evaluating the systems: PeerPanel — a multi-agent scientific-review system built to be measured: planted defects scored by a committed rule with its own negative control, per-call run conditions on every record, and the result published as the record shows it — a null.

📄 Research frameworks (manuscripts in preparation):

• Ensemble NeuralODE GRNs in primary AML — NeuralODE training, ChIP validation and influence pipelines for regulatory-network inference
• Nanopore HBV–HCC integration framework — 29-module long-read pipeline for viral integration, structural variants and methylation

🧬 Research line: deep-learning models of 3D chromatin (C.Origami extensions) + in-silico genetic screens across 155 B-ALL patient samples.

🔧 Open source: long-time maintainer and #2 contributor (150+ commits) of NYU-BFX/hic-bench — Hi-C and HiChIP analysis pipelines used across NYU labs.

⚙️ How I build: agentic engineering as daily practice — Claude Code as the driver, worktree-isolated parallel sessions with merge gates, and eval-driven development: work is reviewed by independent fresh-context evaluators and re-evaluated until it converges. What that discipline produces is public — an append-only decision log, a reproduction campaign on public GEO cohorts whose results include a failed immunoprecipitation in published data that the original depth-only QC could not have seen, and the upstream defects it surfaced. My day-to-day project management runs on an open-source, local-first agentic second brain (someone else's framework — persistent memory, deterministic hooks, background workers) that I run daily and extend with my own skills and tooling. Deterministic code gathers, validates and writes; the model reasons only where judgment is genuine.

Shared first author, Molecular Cell (2025) · Publications · LinkedIn