/AI-Agents
AKIS
Verifiable AI software-development platform.
Open source (Apache-2.0), self-hostable
- TypeScript
- Node.js
- React 19
- PostgreSQL/pgvector
- Docker
- Multi-LLM
- Ed25519
- MCP
The problem
AI coding agents routinely report success without proof — "false green" — and can push unreviewed changes straight into real systems.
What I did
Built a multi-agent platform where every build passes four structural gates (spec approval, verify, push confirm, external write) and can only be marked done once a real boot-smoke test passes.
Highlights
- Multi-agent engine — Scribe, Proto, Critic and Trace — orchestrated behind 4 structural human-approval gates.
- Nothing ships until a human approves and a real boot-smoke test proves the build actually runs.
- Ed25519 build provenance and MCP-based approval gates for every external write.