Neelabh Kumar
Builds AI agents, from model to infrastructure.
now
september 2026writingTraining a small model to run a house
I trained Qwen3-1.7B, an open model small enough to run on a consumer GPU, to control a house through Home Assistant. On the Home Assistant community's voice-command test it went from 75.0% to 89.8%.
75.0to89.8
on the community's voice-command test, before and after training.
PLATES I—VIII
hua-1.7b
A 1.7B model trained to run a house through Home Assistant.
Computer use
A local 8B computer-use model on 50 OSWorld tasks.
Postern
One gateway between your agents and your data, on hardware you own.
ClawLens
Observability and guardrails for OpenClaw agents.
JitAPI
An MCP server that finds only the endpoints a task needs.
PunkClanker
An agent fleet that ran WhatToStream on its own.
Apple Health Bot
Ask your Apple Health data questions in plain English.
Morsand
A 149-page manuscript about my ancestral village, digitized.
upstream
- sep 2026openA row for hua-1.7b on the leaderboard, on all three test sets.allenporter/home-assistant-datasets #343
- sep 2026openFour task checkers passed without the task being done. They now check the result.xlang-ai/OSWorld #593
- sep 2026mergedThe harness dropped code from replies cut off before the closing fence. Now it keeps it.xlang-ai/OSWorld #582
- sep 2026openA failed setup step was logged as complete and the task ran anyway. Now setup failures fail the task.xlang-ai/OSWorld #592
- aug 2026openThe task detail page crashed on timestamps with microseconds. Fixed.xlang-ai/OSWorld #581
- apr 2026mergedApproval callbacks in the Telegram plugin shared a lane with the agent turn that was waiting on them. Gave the callbacks their own lane.openclaw/openclaw #64979