Writing
Mostly on Medium — building things end to end, and what the numbers actually said.
— LLM serving: building the infrastructure layer around LLM inference, one project at a time —
Rebuilding the layer around vLLM, on a laptop with no GPU
A request queue is admission control, not batching. One backend environment variable took throughput from 105 to 218 tok/s and at-capacity time-to-first-token from 2.26 s to 0.31 s; the gateway's own code never changed.
— Gurukul: a series on building a robot learning stack —
Gurukul: Curating Robot Demonstrations with Ray
78,000 public episodes scored with Ray, public videos decoded in an isolated container; curated data beat random, 15 of 20 against 8 of 20.
Gurukul: Serving and Observing a Robot Policy
A policy server, Prometheus and Grafana, and a slow server on purpose — 9 of 20 served.
Gurukul: Teaching a Robot Arm by Imitation, from Scratch
45 demonstrations, an ACT policy, and 10 of 20 on the real arm — most failures trace back to data.
Trust, then verify: a guardrailed LLM agent for edge AI fleets
The agent is the easy part: a deny-by-default policy gate, verification, rollback, and an eval the model can't game.
Eudexia — An LLM-based AI Assistant
A RAG assistant over private docs: a 30-minute hunt through internal documentation, down to under a minute.
— A three-part series on building a self-optimizing 5G network —
Network Observability Framework
Prometheus, Thanos, and Loki feeding Kubernetes logs and metrics to an RL agent in near real time.
Self-Perfecting Networks
A reinforcement-learning agent as the brain of a 5G core, trading service quality against cost in a closed loop.
Open Source Machine Learning Architecture for Telco
A modular ML framework from ingestion to deployment, with swappable compute backends.