Build a decision app, an API or an embedded card by describing it in plain English — over your own database. When an expert overrides a recommendation, it asks why and applies that judgement to the next case. Self-hosted, Apache-2.0, runs on open models.
AI stalls on the calls that carry money — claims, loan sanctioning, clearing a substation — because what decides a hard case isn't in your SOPs or your tables. It's in one officer's head at the moment they decide. Citra captures it there: it recommends, they override, it asks why. Three agreeing makes a named, reversible rule the next case is decided with. Describe it in plain English and get an app, an API or an embedded card over your own database. Apache-2.0 core, self-hosted, open models.
Citra AITransform Enterprise Operations with Sovereign AI
Launched on April 27th, 2026
Reviews
No reviews yetBe the first to leave a review for Citra AI
Maker
📌
Hi Product Hunt 👋 I'm Rohit.
We built Citra for the decisions that carry real money — approving a loan, settling a claim, releasing an aircraft part, clearing a substation. The ones where an experienced person still has to look.
The problem. A loan file passes every automated check: payslips, bank statements, good credit score. Six months later it defaults — the employer never existed. An experienced officer would have caught it, not from the file, but because they know sourcing agents get paid when a loan is disbursed, not when it's repaid. So on those files, they just phone the employer.
That knowledge isn't in your data. There's a field for whether an income document is present — none for whether it's true. And it isn't in your policy either. It's in one person's head, and it leaves when they do.
What we built. Describe your operation in plain English and Citra builds a working app over your own database — or an API, or a card embedded in the screen your team already uses. Every case arrives analysed, scored and recommended, with the policy passage it cites.
Then the important bit: when your expert overrides it, it asks why. When three different experts make the same correction, it becomes a rule — named, attributed, and switchable off — and the next matching case is decided with it.
We tested honestly, and published the failure. We seeded four rules and ran the same files twice, each rule on and off. Three did nothing at all — they restated policy the system could already read. The fourth raised a verification check on 14 of 19 files versus 1 without it. We publish the three failures because they're what make the fourth believable.
Apache-2.0, self-hosted, runs on open models — your data never leaves your infrastructure. visit at https://github.com/Trustedwear-T...
I'd love to hear where you'd point this in your own operation. Happy to answer anything.