Self-built product
Policy Desk
A compliance RAG chatbot that quotes the exact policy clause behind every answer, says so when the policies don’t cover a question, and refuses requests to get around a control.
A complete product I designed, built and tested myself. Client work is in Case studies.

A compliance team can’t use a chatbot that guesses
Compliance and payments staff need answers they can act on and defend. A chatbot that improvises, or quotes a rule that isn’t in the policy, is worse than no chatbot at all, because someone will act on it.
Every answer quotes its clause, and the gaps are honest
Every answer quotes the policy and cites the exact clause, which opens in one click. When the policies don’t cover a question, Policy Desk says so and shows the closest sections instead of inventing an answer. Requests to get around a control, like splitting a payment to stay under a threshold, are refused with the clause that forbids it.
A payment pre-check applies 13 rules from the payments, sanctions and anti-money-laundering policies as the operator types: sanctioned countries, approval limits, politically exposed persons, new clients and split payments. Each finding links to its clause, and checks save to an exportable case log. Teams can upload their own PDFs and ask about them at once, and a trace shows how each answer was found.
Built to run with or without an AI key
Policy Desk is a React and TypeScript application. It runs without an AI key, answering by quoting the policies directly. With OpenAI connected, it writes plain-language answers and removes any citation the model invents, so every reference still points at a real clause.
Tech stack
- React
- TypeScript
- Optional OpenAI
Measured on questions it was never tuned on
On questions it was never tuned on, Policy Desk passed 14 of 17 and found the right clause in the top three 93% of the time.
Screens from Policy Desk

Payment pre-check: each finding links to the clause it comes from.

A question the policies don’t cover, and a refused request to get around a control.

The trace behind an answer: how it was found.

The policy library, with a document the team uploaded.

The evaluation on questions it was never tuned on.
Want something like this for your business?
Tell me what you need it to do. I’ll reply within 24 hours on weekdays with questions or a suggested first step.
Start with the $150 scoping session, credited toward the build.