R&D
Rosicodex
A provenance instrument for AI-assisted science. Every AI answer ships with a receipt a reviewer can verify.
The problem
When an AI tool reads instrument data and returns a fluent answer, a reviewer has no way to reproduce it or check how confident it should be.
What I built
- An agent over a mock analytical lab (Mass-Spec, HPLC, XRD, NMR, SEM) that emits a content-addressed receipt for every answer.
- Each receipt carries a SHA-256 hash of the full prompt and the exact source records the model saw, chained into a tamper-evident ledger.
- Out-of-domain questions and prompt-injection attempts are refused before the model is called, and the receipt records why.
- An evaluation suite gates every change.
The interesting decision
The model never grades itself
Confidence is a deterministic function of the retrieval and the query: coverage, recency, volume, and how specifically the question was parsed. Every input is recorded, so anyone can recompute the score from scratch.
Want something like this?
Book a 20-minute call