What I'm building, contributions to open source, and projects for fun.
Contributing to the open-source community — merged PRs and collaborations on larger repos.
Open-source contributions to NVIDIA's LLM RL stack.
Contributor to NVIDIA's NeMo RL — a scalable library for reinforcement learning and post-training of large language models.
My own projects — research, tools, and things I build for fun.
An open-source self-check harness for distributed training failures.
A chaos lab for PyTorch distributed training: launch a real training workload, inject controlled stressors, and collect evidence from DCGM/NVML, NCCL, XID, thermals, power, PCIe/NVLink, memory pressure, and correctness checks. DCGM is treated as a low-level health data source, not the product boundary; the goal is reproducible failure scenarios, comparable run reports, and actionable preflight/triage signals for GPU clusters.
A governance substrate for two coding agents in one repo.
Lets Claude Code and Codex work in the same project safely, coordinating through the filesystem instead of the token stream: capability policies, advisory locks and a shared task board, worktree isolation with merge gates, and an append-only journal for replayable oversight. Designed around three axes — security, controllability, and cost. A single dependency-free Go binary.
A bilingual, 100% client-side wealth-projection simulator (v1.0).
Project net worth year by year (nominal vs. inflation-adjusted) from careers with equity/RSU, a home/mortgage what-if, and market assumptions — with a Monte Carlo mode (p10/p50/p90 + success probability), stress tests, FIRE, and goal-seek ("$X by age Y"). Bilingual EN / 中文, installable offline PWA. 100% client-side — no backend or tracking; open-source (MIT).