I build the evidence layer for AI — and I test it on a 36-year-old company before anyone else has to trust it.
A self-hosted layer between AI assistants and real data: deterministic PII masking, policy enforcement, row caps, hash-chained audit and Ed25519 receipts anyone can verify offline. MCP-native, open source. Born inside my own ERP — 121,366 real identities pseudonymized, zero leaked.
conarium.dev →When agents start spending money, the hard question is not the payment — it is the evidence. Cedulon signs the offer before the spend, denies by default, signs the receipt after, and reconciles both against the payment rail, so a settlement with no receipt is caught. Missing evidence is itself evidence. Specified in public IETF Internet-Drafts; eight packages on npm with build provenance.
cedulon.com →A memory format for AI agents where every claim carries its source, its age and its boundary — the subject the agent must not improvise on. Stale facts announce themselves instead of ageing quietly into confident answers. Apache-2.0, free, served over MCP. Formerly Talamus.
tugra-ai.com →Paste any product URL and see the item on yourself before you buy — no fitting room, no returns. iOS, Android and web. The consumer end of the same company: retail is where I learned what people will actually tolerate from an AI.
wearu.app →tugra ·
@cedulon/* (eight packages) ·
@conarium-ai/core
Published to npm with SLSA build provenance — you can check who built the bytes you install, not just who says they did.
A 36-year-old İzmir furniture brand with 5 stores, rebuilt AI-first around a system that reads 100,000+ customer records and 10,000+ incoming messages a month. It grew +102% year-over-year. It is not a case study — it is where every product earns the right to exist, and where I find out what breaks before a customer does.