Project Kintsugi.
Self-evolving AI for mutual aid networks. Runtime constraints authored as VALUES.json. The breaks in continuity, mended with care, become the strongest parts.
The core innovation
Safe self-modification. Most AI systems are trained once and deployed forever. Kintsugi can evolve its own decision weights based on real-world outcomes, fork shadow copies to test changes before applying them, detect value drift and auto-correct toward its ethical baseline, and learn from stakeholder feedback while requiring consensus for major changes. All within guardrails. The system literally cannot evolve away from its core values.
Why community organisations
Most AI infrastructure assumes profit-driven deployment. Kintsugi assumes you are coordinating a mutual aid network, managing nonprofit grants, practicing democratic governance in a worker cooperative, or protecting vulnerable communities. Every architectural decision reflects these values. The system is designed to serve communities, not extract from them.
VALUES.json
Organisations define their ethical constraints in a small JSON file the runtime loads on every decision. Beliefs like community_first, profit_seeking:false, transparency_default. Constraints like never_share_pii_externally, require_consent_for_data_use, prioritize_vulnerable_populations. These are not suggestions — they are hard constraints that self-modification cannot override.
Capabilities.
| capability | what it does |
|---|---|
| Memory | 3-stage pipeline: extraction → significance scoring → hybrid retrieval, with temporal decay and per-org isolation |
| Skill chips | 22 domain handlers across core ops, programs & people, and community aid |
| Multi-platform | Slack, Discord, web chat, email — all share one agent loop |
| Deployment tiers | SEED (laptop) · SPROUT · GROVE · FOREST (full cluster) |
| Multi-tenancy | Row-level, schema, or full database isolation per org |
Roadmap.
This research is ongoing. If your work intersects with ours —