Projects
These open-source projects are where I test the same ideas I bring into consulting work: AI-agent control boundaries, retrieval architecture, multimodal search, secure tool use, and product workflows that can move beyond a demo.
The point is not novelty for its own sake. Each project is a working system that exposes real design decisions: what the model is allowed to do, where authoritative data lives, how users inspect results, and what has to exist around the AI layer for the product to be useful.
What this shows
I use these projects as applied research, not showpieces. They let me pressure-test architectural patterns against real code, real data, real dependencies, and real operational constraints.
For clients, that matters because AI strategy becomes useful only when it survives implementation. The same questions keep coming back: how should agents use tools, how should specialist data be retrieved, how should risky actions be controlled, and how should the surrounding product make AI outputs inspectable?
These projects are small enough to understand and concrete enough to prove the judgment behind the advice.
Case studies
01
Charmant.AI is an open-source local AI agent that explores how terminal agents can use real tools safely: explicit permission modes, dynamic shell-command risk classification, reusable skills, project instructions, and auditable execution.
02
OCRE AI is an open-source applied-AI system built on 43,000+ documented Roman coin types. It combines structured catalog browsing, hybrid vector and full-text search, image-based identification, and an AI chat agent grounded in real numismatic data.
Next step
I help teams make the architectural decisions around AI systems that become expensive to fix later: retrieval design, agent permissions, AWS security posture, governance, observability, and delivery readiness.