Software Engineer — Applied AI & LLM Infrastructure. 14 years of production systems (ex-AWS S3), building agent tooling in public: the DSH plugin suite, model routing, autonomous orchestration. Insurance software background (Segfy, Cotak). One rule underneath everything: the shortest working diff wins.
One command installs it all:
dsh plugin --profile <profile> add github:andrepontesmelo/dsh-suiteThat single command installs the bundle, pulls both plugin dependencies (@andrepontesmelo/dsh-model-router, deep-horizon), and points DSH skill discovery at the nine bundled skills.
| Repo | Type | What it does |
|---|---|---|
dsh-suite |
bundle | The entrypoint: 9 agent skills + 2 plugins, installable in one command |
dsh-model-router |
plugin | Virtual model ids routed over real provider/model candidates — priority failover with exponential backoff, round-robin rotation, sleep windows |
deep-horizon |
plugin + CLI | Keeps a project's aim — one about line, up to 5 gaps — in front of every agent session, via per-harness adapters over one shared .horizon store |
dsh-wayfinder-ui |
plugin | Spawns one live session per ready Wayfinder task and renders the dependency map in the DSH chat GUI |
archloop |
standalone | Overnight architecture-improvement loop driver — auto-scan + refactor cycle for git repos |
hkrc |
standalone | Hermes Kanban Recovery Controller — portable, instance-scoped blocker recovery |
Also public, kept out of the bundle: moving-target (plugin, deprecated — superseded by deep-horizon), and two forks carrying local features — dsh-telegram-channel (Telegram mobile remote: /new, gesture passthrough) and deepseek-harness-mobile (Android companion). Legacy: publications (2016, archived).
I design agent orchestration and self-improvement loops that keep code clean, scalable, and maintainable in the AI era — context and memory management, multi-agent workflows, intelligent model routing — built and tested in public here, and run daily in my own workspace.