Neelabh Kumar Staff engineer at Broadcom with about ten years building infrastructure and products: VMware Cloud Foundation since 2020, Oracle Cloud Infrastructure before that. I usually start with a problem I run into myself. Outside work I build what agents run on and train small models with RL: - Hua (hua-1.7b): Qwen3-1.7B trained with GRPO to control a house through Home Assistant. 75.0% to 89.8% on assist-mini, trained on one RTX 3060. - Postern: a self-hosted personal context gateway. Connect your mail, money, calendar, health and home once, and any agent you allow reaches them over MCP, deny by default. - ClawLens: observability and guardrails for OpenClaw agents. It led to a merged upstream fix for an OpenClaw approval deadlock (300-second timeouts down to about 3 seconds). - OSWorld: a merged parser fix and open PRs tightening evaluators that could pass unfinished tasks. What I care about is making agents useful and reliable: giving them the right context, controlling what they can do, and measuring whether they did the job. Rooted cosmopolitan. Off the keyboard: hiking, running, 3D printing, automating my home.
Neelabh Kumar Staff engineer at Broadcom with about ten years building infrastructure and products: VMware Cloud Foundation since 2020, Oracle Cloud Infrastructure before that. I usually start with a problem I run into myself. Outside work I build what agents run on and train small models with RL: - Hua (hua-1.7b): Qwen3-1.7B trained with GRPO to control a house through Home Assistant. 75.0% to 89.8% on assist-mini, trained on one RTX 3060. - Postern: a self-hosted personal context gateway. Connect your mail, money, calendar, health and home once, and any agent you allow reaches them over MCP, deny by default. - ClawLens: observability and guardrails for OpenClaw agents. It led to a merged upstream fix for an OpenClaw approval deadlock (300-second timeouts down to about 3 seconds). - OSWorld: a merged parser fix and open PRs tightening evaluators that could pass unfinished tasks. What I care about is making agents useful and reliable: giving them the right context, controlling what they can do, and measuring whether they did the job. Rooted cosmopolitan. Off the keyboard: hiking, running, 3D printing, automating my home.
Nov 2025 — PresentPalo Alto, CA
AI agents for VMware Cloud Foundation operations. Built an incident-triage tool on-call engineers adopted without a mandate; it reuses only diagnoses confirmed by root-cause analysis. Built a known-issue triage agent that checks every fix and KB citation against source runbooks.
Oct 2023 — Nov 2025Palo Alto, CA
Infrastructure lifecycle for VMware Cloud Foundation, continuing after Broadcom acquired VMware.
Jun 2020 — Oct 2023Palo Alto, CA
Core contributor to the VCF Async Patch Tool, used by thousands of enterprise customers. Owned CI/CD, telemetry and failure recovery, and designed provisioning of pre-patched infrastructure domains that removed multi-day maintenance windows.
May 2019 — Aug 2019Atlanta, GA
End-to-end Zigbee integration for VMware Pulse IoT Center: device discovery, automated onboarding, telemetry and remote commands.
Oct 2015 — May 2018Bengaluru, India
Oracle Cloud Infrastructure and Oracle HTTP Server: infrastructure software and release tooling in Java and Python. Built automated build-quality checks that supported weekly releases.
Aug 2014 — Oct 2015Bengaluru, India
Automotive communication software for Jaguar Land Rover in embedded C over CAN/LIN. Started an internal technical seminar series.
Hua (hua-1.7b): Qwen3-1.7B trained with GRPO to run a house through Home Assistant, rewarded only by pass/fail from a real Home Assistant run. 75.0% to 89.8% on assist-mini, on one RTX 3060.
Self-hosted personal context gateway for AI agents. Nine integrations across mail, money, calendar, health, contacts and home; one key per agent, deny by default, every call logged.
Observability and guardrails for OpenClaw agents: risk-scores every shell command in under 5 ms, blocks or requires approval by rule, and keeps a hash-chained audit log.
Natural-language streaming search: say what you want to watch, get a shortlist with watch links. Its development and growth were run by a fleet of OpenClaw agents with capped permissions; it reached thousands of monthly users at its peak.
