Production AI systems, shown by what they do. Built over 25 years in software, including running my own company end-to-end. Each one has its own repo below, with the design, runnable code, and clips of the real thing. These run on real money and private data, so the public repos show the real architecture, runnable samples, and live clips, sanitized where the internals stay private.
Proving Ground · ⭐ Showcase
Generative and retrieval AI over vast game data. A contract-blind LLM drafts warships from one sentence; a grounded assistant reads 149,288,931 rows of synthetic combat telemetry in indexed Postgres and cites the exact rows or refuses. One typed contract judges every row that moves.
Concierge · 🟢 Production
Drafts grounded replies to customer emails over typed MCP tools, into Gmail Drafts behind a policy gate and human review. Never auto-sends; prompt-injection attempts are quarantined.
Quant Watchtower · 🟢 Production
Read-only MCP operations console over a live 24/7 algorithmic trading fleet. Sanitization is enforced in the data layer, so the strategy never reaches the model.
Cortex · 🟢 Production
Self-hosted knowledge-AI platform: an Obsidian vault wired for AI agents through an MCP layer, Docker sync, and Python tooling. Ships a runnable clean-room linter.
Dev System · ⚙️ Method
The method behind all of these: bounded efforts, standing invariants, a per-change audit trail, acceptance discipline. Validated on a live production system for two-plus months.
Celestia · 🟢 In daily production
Persona-driven assistant over a private Obsidian vault through a typed MCP filesystem core. Reads before it answers; every write routes to its owning note.
Ledger · 🟢 Mirrors production
Accounting-ops agent that closes the books over typed MCP tools. Auto-books only what a deterministic policy gate proves safe, routes the rest to a human, prints a full audit trail.
The Librarian · 🔵 Lab
Agentic RAG over a podcast archive: hybrid BM25 + dense retrieval (RRF), answers cited to the minute, scored against a hand-labeled gold set (recall@k, MRR).
Different domains, one architecture: agents get typed, allowlisted tools instead of raw access; a policy gate lives in code, not in a prompt; state changes are reconciled against a source of truth; and anything that leaves the system passes a human. Building the same shape over many kinds of data is the point. The pattern and the discipline behind it are documented in Dev System.