Builds AI agents, from model to infrastructure.
Staff Engineer at Broadcom (VMware Cloud Foundation), San Francisco. On my own time I train small models with RL and build what agents run on: context, permissions, guardrails, evals.
TypeScript · Python · MCP · evals · RL (GRPO, LoRA) · computer use · neelabhbuilds.com
| Project | What it is | Links |
|---|---|---|
| hua-1.7b | Qwen3-1.7B trained with GRPO to run a house through Home Assistant. The only reward was pass or fail from a real Home Assistant run, no supervised examples. On the community's voice-command tests it went from 75.0 to 89.8 on assist-mini and from 51.5 to 68.0 on assist, with questions unchanged. Seven hours on one RTX 3060. | Write-up · Model · Serving code |
| Postern | Self-hosted personal context gateway for AI agents. Connect mail, calendar, money, health, contacts and home once on your own machine. Any agent you allow reaches them over MCP with a key and a grant. Agents never see a provider credential. OAuth 2.1 server written from scratch, deny-by-default grants, append-only audit log. Apache-2.0. | getpostern.com · GitHub |
| ClawLens | Observability and guardrails for OpenClaw agents. An in-process hook parses every shell command and scores it 0 to 100 in under 5 ms with no model call. Rules block, require approval or notify. Every decision lands in a hash-chained audit log. | Project page · GitHub · npm · ClawHub |
| Computer use / OSWorld | Harness and evaluator fixes to the desktop computer-use benchmark. Ran 50 tasks with a local 8B computer-use model and read every failure against the checker and the action log. 17 of 24 failures were the model claiming success on a task it had not done. Three PRs merged upstream. | Notes · OSWorld |
| JitAPI | MCP server for just-in-time API discovery from any OpenAPI spec. A dependency graph plus local embeddings return the endpoints a task needs, in call order. Against GitHub's 800-endpoint spec loaded whole: 34x fewer input tokens, 1.4x faster. | GitHub · PyPI · Project page |
| WhatToStream.ai | Consumer streaming search, 2025–26. Four OpenClaw agents ran the site, each with its own Telegram bot and a cap on what it could change. Retired September 2026. | Project page |
| Morsand | The history of my ancestral village, recovered by machine from typewritten English and handwritten Devanagari. | morsand.com |
Also: Apple Health Bot, questions over your Apple Health export in plain English · HA agent, a Home Assistant agent over Telegram with self-deleting automations.
- openclaw/openclaw #64979, merged: a deadlock in the Telegram approval flow. Approvals that timed out at 300 seconds resolve in about 3.
- xlang-ai/OSWorld #582, merged: the harness dropped code from replies cut off before the closing fence.
- xlang-ai/OSWorld #593, merged via #622: four task checkers passed without the task being done.
- xlang-ai/OSWorld #581, merged: the task detail page crashed on timestamps with microseconds.
- allenporter/home-assistant-datasets #343, open: a row for hua-1.7b on the leaderboard.
