I learn a field by building something real in it. I start with the people doing the work, find where time or quality is lost, build a first version and keep improving it where it's used. Most of what's here came out of that, across software, CAD, sensors and model training.
| Project | What it is |
|---|---|
| lume-enclosure | Parametric OpenSCAD enclosure for a depth-camera stage unit. 18 printable parts, joint coupons for fit tests before the big prints. |
| jeli-device | Printable N'Ko learning device: generated body STLs, an adjustable Pi dock with geometry tests, and a browser device-twin. |
| motionmix-core | Rust core of a multi-camera capture hub: Record All as a transaction, a first-frame watchdog, receipts, pose signals. |
| meshd | Rust daemon that runs AI coding agents in terminal slots behind an HTTP API, across machines. |
| thunder-train | Distributed LoRA fine-tuning across two Apple Silicon Macs over Thunderbolt with MLX. |
Some of my work stays private because it holds customer data or details of my own setup:
- KOATJI field app and delivery planner: I sold to cafés in several US markets and built the tools the team ran on.
- MotionMix hub: the full multi-camera server.
- Jeli models: N'Ko OCR, speech and tone models.
The portfolio has drawings, photos and results.
Swift and SwiftUI, Rust, Python, TypeScript · Postgres / Supabase · OpenSCAD, Blender, 3D printing · Raspberry Pi, ESP32, depth cameras · PyTorch, MLX