Building the platforms & infrastructure AI runs on
I work in the unglamorous middle layer of AI, turning research and prototypes into systems teams can depend on. 0→1 is my default: ML platforms, agentic and GenAI infrastructure, and the whole path to production.
The part I love is the bridge - making research-stage ideas hold up and scale, including the bits people skip: governance, cost, adoption.
Outside that: a bike, a race, travel, or a new language (not the programming kind :-)).
Building ML/AI platforms and agentic + GenAI infrastructure: from RL-driven self-optimizing 5G networks to an enterprise Data & AI platform.
Audience-modeling R&D for Advertising Cloud: end-to-end ML for ad targeting at scale.
Data science for T-Mobile: NLP call-driver classification and BI analytics pipelines.
Master's in Data Science & Business Intelligence. TA for ML & econometrics (IMT 574).
R&D in image processing and augmented / mixed-reality applications.
Teaching a low-cost SO-101 arm to pick up a toy and drop it in a bowl by imitation - an ACT policy trained on 45 demonstrations, scored honestly at 10 of 20 on real hardware, with every failure traced back to data. Part 1 of a series on building a robot-learning stack from scratch.
A RAG assistant that lets teams query their own private documentation in plain language and get answers grounded in real sources — collapsing a 30-minute hunt across hundreds of internal docs into under a minute.
— A three-part series on building a self-optimizing 5G network —
The data layer underneath: an open-source, cloud-agnostic observability stack (Prometheus, Thanos, Loki) that collects logs and metrics from a Kubernetes cluster and feeds them to the RL agent in near real time.
How a reinforcement-learning agent can act as the brain of a 5G network that tunes itself — a closed-loop system where the agent observes a simulated 5G core running as a Kubernetes app and adjusts resources to balance service quality against cost.
A modular, open-source ML framework spanning the full lifecycle — ingestion, data prep, training, deployment — built so teams can swap compute backends without re-plumbing. Applied to anomaly detection on Kubernetes performance logs.
A guardrailed LLM agent that monitors a fleet of edge AI inference nodes, diagnoses faults, and remediates them only within hard policy limits. Ships an evaluation harness and a live Prometheus + Grafana observability stack.
An OpenAI-compatible LLM inference gateway — async request queue, SSE streaming, auth, rate limiting, prompt-based routing, OpenTelemetry tracing, and a provisioned Prometheus/Grafana board. Ollama-backed locally, swappable to vLLM/TGI.
A distributed multimodal pipeline on Ray that ingests autonomous-driving sensor data (camera + lidar + metadata), preprocesses it, and runs pretrained-model batch inference at scale — writing structured Parquet. An exercise in ML-infra orchestration: stateless Ray Data tasks for preprocessing, an actor pool for inference.
Teaching a real SO-ARM101 robot arm to pick up a toy and drop it in a bowl - an ACT policy trained on 45 teleoperated demos with Hugging Face LeRobot, scored honestly at 10/20 on hardware. Project 1 of a series working toward the full physical-AI loop: data collection, training, and evaluation done; serving and observability next. (The leader arm is guru, the follower shishya.)