WebdexThe front door
A LibreChat deployment is seven containers, a 38-line config file, two bind mounts and one shadowed upstream file: the chat application is the smallest part of it.
Blog
Notes on software delivery, architecture and AI-augmented engineering.
Written from client work rather than from the sidelines: how delivery actually runs, the architecture decisions that mattered, and where AI-augmented engineering changed the work — including the parts that did not pay off.
One endpoint, many providers
A self-hosted OpenAI-compatible 9router endpoint in front of two flat-rate inference plans: five routes, one key per device, and six patch layers that decide whether it works.
A shared OpenCode configuration
One git repository for the agents, skills, commands and permissions a team runs on, plus two small plugins: one keeps every checkout current, one keeps MCP tool definitions per repository.
Evaluating RAG systems
A retrieval-augmented generation demo passes its own review and still fails in use. The failing set, the metrics and the CI gates that catch it.
MCP servers in production
A demo answers whether a model can call a function. A service has users, uptime and an audit trail: trust boundaries, tool contracts, operability, rollout.
Micro Mono Modu
An architecture choice is a bet on years of running cost: the stages a system passes through, five shapes to pick from, and how to choose between them.