NYU · New York
ML systems, reliability
and reproducible evaluation.
I build and debug the infrastructure around machine learning. My focus is simple: understand the failure, test the fix, and make the evidence reviewable.
From model execution
to developer tooling.
01 / Selected work
Focused projects with inspectable code,
clear boundaries, and questions still to answer.
Make ML runs easier to review.
An early ML infrastructure project for supervised fine-tuning on customer-operated infrastructure, failure recovery, and portable signed execution receipts.
The demo uses synthetic execution with real signature verification. Customer validation and production qualification remain open.
Inspect the workflows around AI agents.
An experimental TypeScript static analyzer for GitHub Actions workflows that run AI coding agents. It examines permissions and paths from untrusted inputs.
Findings support human review. Detection accuracy has not yet been measured.
02 / Open source
Changes merged into projects
maintained by other people.
Fixed torch.compile failures in DETR-family position embeddings and added a compiled-dtype regression test.
Restored interpreter completion fallback when Jedi static analysis raises, with regression tests for failure paths.
Tuya startup retries on network failures, LG ThinQ service-error reporting, and LinkPlay discovery behavior for known devices.
03 / Research
ML inference, evaluation,
and reproducible systems research.
I’m interested in how we evaluate inference changes against both performance and output-quality requirements. I am exploring Fall 2027 PhD opportunities in ML systems.
Earlier work at NYU WIRELESS connected material-aware 3D reconstruction with wireless simulation. I contributed to scene reconstruction, material labeling, and point-cloud processing.
Research & Engineering Profile PDF ↓
Public experiment: TraceBench — serving workload and measurement infrastructure with scoped experimental evidence. Its README defines the experimental scope and reproduction steps.
04 / Experience
Developing ML execution and evidence tooling: job contracts, runner coordination, failure recovery, and release evidence.
Current projectBuilt ComfyUI-based multimodal generation workflows and benchmarked open-weight models for internal research and demos.
May — Aug 2025Supported labs and office hours, helped develop assignments, and contributed instructional material.
Feb — May 2025Worked on Java and Spring Boot monitoring services, multithreading, Redis caching, and MySQL queries.
Jan — Dec 202305 / Field notes
The job boundary, recovery, and what the demo can show.
DESIGN NOTE · AXIOMArtifact identity, missing context, and the limits of signatures.
SYSTEMS NOTEWhat market-data experiments teach about event ordering.
Let’s compare notes
ML systems research, infrastructure engineering, or a workflow CloudTune could help evaluate — I’d like to hear about it.
This opens your email app. No form or account required.