Founder at VERAX. I work on the boundary between AI systems and the things a company cannot afford to get wrong — its data and its money — and on making what happens at that boundary provable rather than asserted.
Three products, one claim: a machine action should carry its own evidence.
conarium.dev · source (MIT) · @conarium-ai/core on npm
A governed gateway between AI assistants and databases. Three things it does, in the order they matter:
- Enforces before disclosure — table allow-lists, row caps, deterministic masking, and columns that may not appear in a predicate at all.
- Signs a portable receipt for every access — hash-chained, verifiable offline by a single file that imports nothing from the package it checks.
- Reconciles against the source's own counters — because a log written by the gateway cannot reveal an access that never reached it.
Running in production at one company: 121,366 identities masked across 121,374 records. What that number is, and what it is not.
cedulon.com · source (Apache-2.0) · @cedulon/* on npm, eight packages with build provenance
When agents start spending money, the hard part is not the payment. It is the evidence.
- A trade manifest is signed before the spend, so the offer cannot be rewritten after it.
- The policy decision point denies by default; an allowed payment gets a COSE_Sign1 receipt.
- Epoch checkpoints and a transparency log make the receipt chain append-only.
- A rail-extract reconciliation compares the chain against the payment rail in both directions, so a settlement with no receipt is named rather than assumed.
Missing evidence is itself evidence. It settles on a mock rail: no wallet is held and no transaction is signed.
tugra-ai.com · source (Apache-2.0) · tugra on npm
A memory format where every claim carries its source, its age and its boundary — the subject the agent must not improvise on, which stays quarantined until a human approves it. Stale facts announce themselves instead of ageing quietly into confident answers.
Four IETF Internet-Drafts, sole author. Individual submissions: not adopted by a working group and not endorsed by the IETF. The drafts, the review threads and the interoperability results are public, which is the point.
| Draft | Rev | What it defines |
|---|---|---|
| draft-dogru-cedulon | -05 | The Cedulon protocol: manifest, decision token, receipt, checkpoint, reconciliation |
| draft-dogru-cedulon-reattestation | -00 | Carrying spend evidence across algorithm retirement |
| draft-dogru-cedulon-streaming | -00 | Continuous completeness for agent spend |
| draft-dogru-scitt-disclosure-evidence | -07 | Transformation evidence and coverage reconciliation, as Signed Statements on a SCITT Transparency Service (RFC 9943) |
The SCITT draft deliberately defines payloads only. Delegation receipts, policy replay and transparency logs are other people's work, and the draft says so.
A vendor-neutral adversarial conformance suite for governed data access. Any implementation can run it through a small adapter; the runner imports nothing from mine. It reports capability status rather than a score, and it ships the cases my own implementation fails next to the ones it passes — a suite whose author scores full marks proves nothing.
Claims are cheap, so these projects publish their own limits: LIMITATIONS.md lists what Conarium does not do, including the inference channels that remain open. PRIOR-ART.md lists everyone else I could find doing similar work, and exists to be corrected.
Everything ships first inside my own 36-year-old furniture company, where a bad assumption costs real money before it can cost anyone else's.
One-person team. No SOC 2, no independent penetration test, one production deployment. Those are on the roadmap, and they are not claimed before they exist.
📫 e.dogru@conarium.dev · conarium.dev · cedulon.com · tugra-ai.com · emekcandogru.com