Projects, analysis, and explanations of the systems I work with.
The architecture behind Hermes, my homelab agent: a sandboxed tool runner, a guarded API layer and an approval loop that let it provision a Docker-ready Proxmox container from one message.
A Gemini agent searches for movie scenes and songs, shows its tool calls, and returns a YouTube result inside the app.
A Gemini agent connects to MCP tools backed by PostGIS and pgRouting. The interesting part is resolving a place, choosing stations, and showing where the answer came from.
I exposed the blog’s retrieval system as three read-only MCP tools. The implementation is small enough to inspect, and the same search tool appears in recorded agent sessions.
A grounded prediction agent competed under the same scoring rules as people. The published group-stage chart reports 47 correct outcomes from 72 picks; here is the implementation and what that result still needs to establish.
I put masking inside the producer’s data tool, then limited the tools that agent could call. Here is the code, the synthetic example, and the boundary this proof of concept actually provides.
6 of 6 articles
latent