AI educator and systems builder at U of Digital. Builder's notes on useful systems, published weekly with the receipts left in.
A family of open-source GPU servers on Modal, plus one operator CLI that runs the whole fleet. The runbooks are war stories, not checklists.
| Package | What it does |
|---|---|
| modal-toolkit | The operator CLI: init, pull, deploy, doctor, dashboard, cost |
| modal-inference-server | OpenAI-compatible GPU inference with hot-set routing |
| modal-embedding-server | GPU embeddings with a sync protocol and model-selection evals |
| modal-vision-server | Classification and segmentation driven by a model card |
| modal-finetune-server | LoRA fine-tune, merge to GGUF, honest eval gates |
| modal-vault-server | A hosted vault and MCP memory plane for agents |
| modal-victoriametrics-server | Single-node VictoriaMetrics, CPU only, scale to zero |
| modal-shared-libs | Stdlib-first shared modules, vendored into every package |
There is a paid side too: a Modal Production Operator Kit and a set of architecture services, for teams that want the work done with me rather than handed to them.
Shared bash primitives for machine hygiene, an MCP server for capturing and analyzing browser traffic, a model router that picks the right model tier per turn, and a way to run a second coding agent as a persistent service.
- pi-link - a local WebSocket network for coordinating Pi coding-agent terminals
- pi-model-router - an intelligent per-turn model router
- kylebrodeur.xyz - the index of ongoing work
- Builder's Notes - the writing
- PlantFluent - plant identification and a garden journal
- Hugging Face · npm · Ollama
Building notes on useful systems, mostly. The receipts are in the repos.