Systems engineer. Elixir/OTP, Erlang, Python.
Building agent infrastructure that treats LLM orchestration as a distributed systems problem, not a prompt engineering problem.
VAOS (Viable Autonomous Operating System) — an Elixir/OTP framework for epistemic agent orchestration. Three libraries, one runtime:
| Library | What it does | Size | Tests |
|---|---|---|---|
| vas-swarm | Agent orchestrator. Signal classification, 3-tier routing, 48 tools, 18 LLM providers. | 92K lines Elixir + 20K Rust | ~3,320 |
| vaos-ledger | Epistemic engine. Claims, evidence, attacks tracked with Expected Information Gain scoring. | 6,718 lines | 208 |
| vaos-knowledge | Triple store. ETS-backed, SPARQL subset, 4 OWL 2 RL rules. | 1,095 lines | 108 |
The core idea: agents should reason about what they don't know (epistemic governance) rather than just executing tool chains. The investigate tool uses adversarial dual-prompt architecture with citation verification — it argues against its own findings before reporting them.
vaos.sh — SaaS interface to the VAOS runtime.
| Repo | Language | What |
|---|---|---|
| vggt-mps | Python | Apple Silicon (MPS) port of Meta's VGGT visual geometry grounded transformer |
| MCPhoenix | Elixir | Model Context Protocol gateway for Phoenix applications |
| wf-substrate | Erlang | 43 workflow patterns compiled into native OTP control structures |
| speckit | Shell | Spec-driven development toolkit |
| cloudflare_durable_ex | Elixir | Cloudflare Durable Objects client for Elixir |
| ace-playbook | Python | Adaptive Code Evolution — self-modifying agent architecture |
The architecture draws from Stafford Beer's Viable System Model (1972) — recursive autonomy with meta-systemic coordination. Each agent subsystem mirrors VSM's System 1-5 structure: operations, coordination, optimization, intelligence, policy.
References: Beer's Brain of the Firm, Ashby's Law of Requisite Variety, Wiener's cybernetic feedback loops. See the viable-systems org for the upstream VSM implementation.