AI/ML engineer. I train language models and build the agent systems that run on them.
My work sits at the seam between the two: post-training and evaluating LLMs on one side, shipping the orchestration layers, developer tools, and products they run inside on the other. I care about the part most demos skip — real workflows, real users, and enough depth to keep improving after launch.
agent-relay.dev · source · 8,000+ developers · Apache 2.0
The control plane for coding agents. A local-first CLI + desktop app that launches multiple coding agents, streams their work in real time, hands sessions between models without losing context, and resumes anything that stopped mid-task.
- Adapters for Claude Code, Cursor, VS Code, and Warp
- Local-first by design — sessions and context stay on your machine
- Live multi-agent monitoring, mid-flight model handoff, session resume
- Shipped to 8,000+ developers; open source, on PyPI as
agent-relay-tool
curl -fsSL https://agent-relay.dev/install.sh | shAI-native product discovery that turns signal into specs. Ingests customer signal from interviews, surveys, support tickets, and analytics, finds the patterns across thousands of data points, and generates execution-ready PRDs, design proposals, and technical specs. Ships handoff packages built for AI coding tools, and integrates with Slack, Jira, Linear, Figma, GitHub, Notion, and Confluence. Teams using it cut discovery cycles from weeks to hours.
Live disaster intelligence. Mobile app and organizational dashboard combining official hazard feeds with community-reported incidents to deliver address-level risk for fire, flood, storm, and earthquake. Individuals get tunable alerts by radius, hazard, and time; operations and public-sector teams get a filtered, mapped situational view with API/webhook integration and role-based access. Trusted by teams at institutions including Stanford, MIT, Carnegie Mellon, and UC Berkeley.
AI-native interactive film. Turns viewers into participants using multimodal reasoning, long-context memory, and a narrative-state engine that adapts scenes, dialogue, character memory, and story consequences in real time.
Built at the Moss Conversational AI Hackathon at Y Combinator — a voice-first operations layer for grounded internal knowledge retrieval, multi-agent monitoring, and workflow visibility.
Hands-on across the post-training stack, not just the API surface:
- Fine-tuning — SFT, LoRA/QLoRA, and instruction tuning on open-weight models against custom, task-specific datasets
- Post-training & alignment — RLHF, DPO, and reward modeling to shape model behavior past the base checkpoint
- Evaluation — eval harnesses, agent and model benchmarking, and regression suites that catch quality drift before users do
- Training infrastructure — dataset curation and tokenization, distributed training, GPU orchestration, and experiment tracking
The through-line with Agent Relay: models are only as useful as the systems routing work to them, so I build both ends.
AI & ML — PyTorch, TensorFlow, Hugging Face, LoRA/QLoRA, RLHF/DPO, evals & benchmarking, RAG, multimodal AI, voice agents, OpenCV, OCR, Pandas, NumPy, Jupyter
Data & infra — Apache Airflow, Apache Flink, PyFlink, Docker, Kubernetes, Argo, CI/CD, Google Cloud, Firebase
Product — Next.js, React, TypeScript, Angular, Tailwind, Flutter, Dart, Swift, React Native, Expo
Backend & systems — Python, Rust, C/C++, Java, Node.js, Express, Flask, PostgreSQL, MySQL, SQLite, Linux, Bash
Design — Figma, Photoshop, Illustrator, UI/UX