Two active projects — one in AI-powered media, one in autonomous robotics. Both share the same approach: find a real coordination problem nobody has solved, and build an open system to fix it.
C.R.E.W. — Coordinated Robot Emergency Workforce
An open-source protocol that integrates commercial robots into a unified emergency response network — no new hardware required.
The problem: Tens of thousands of commercial robots sit idle during disasters with no way to assist first responders. They have sensors, mobility, and compute — but no coordination layer.
The solution: A software-only protocol built on ROS 2 that lets robots broadcast their capabilities and volunteer for emergency tasks, while keeping humans in control of every assignment.
In this demo: a multi-car pileup at 3:47 PM. Ambulance is 8 minutes out. Three nearby delivery robots receive the broadcast — two respond, one declines because it’s mid-delivery. This is CREW. Robots help when they can. Humans decide.
| ROBOT | TYPE | STATUS | ASSIGNMENT |
|---|---|---|---|
| delivery_drone_01 | Delivery Drone | ✅ Available | Aerial routing for ambulance — ETA 6 min |
| delivery_drone_02 | Delivery Drone | 🔴 Mid-delivery | Declined — continues current task |
| food_delivery_bot_01 | Ground Robot | ✅ Available | Ground damage assessment — ETA 4 min |
CREW works because it respects the real world: not all robots are always available, and no robot is ever commandeered. The protocol matches emergency needs to available assets — then a human coordinator approves every assignment before any robot moves.
Podcast Automation Pipeline
An end-to-end system that takes any audio source — radio shows, YouTube channels, morning broadcasts — and makes every spoken word searchable via natural language.
The problem: Spoken audio is unsearchable. Hours of radio, podcasts, and YouTube content exist with no way to find specific moments, topics, or conversations.
The solution: A fully automated pipeline that downloads, transcribes with speaker diarization, embeds into a vector database, and publishes a searchable archive — running daily on a cron job with no manual steps.
Currently powering three live archives across 19,026 indexed chunks. Type a question in plain English and the system returns the exact moment it was discussed — with speaker labels and a playable audio timestamp.
The pipeline is input-agnostic — it works on RSS feeds, YouTube channels via yt-dlp, or any audio file. Once ingested, the entire archive is queryable in seconds. Cost: approximately $0.05–0.15 per episode in Claude API calls.