IntentKeeper

Inspiration

AI-assisted tools can help people build applications quickly, but a nontechnical owner still needs a clear way to verify that the finished app keeps its original promises.

IntentKeeper was inspired by that gap between generated code and human accountability. It gives an owner an evidence-based checkpoint before release.

What it does

IntentKeeper checks whether an AI-built reservation app still follows three owner-defined safety promises:

1. Customer contact data is never public.
2. Customers can cancel their own reservation.
3. Only an admin can view customer contact details.

The result is either COMPLIANT or NOT COMPLIANT, with the evidence behind each check and a next step chosen by the human owner.

How we built it

IntentKeeper is built with Python, Streamlit, and the Strands Agents SDK.

Three project-owned deterministic tools inspect observable behavior: public PII exposure, customer cancellation, and contact-role access. These tools decide every pass or fail result and return machine-readable evidence.

The Strands explanation layer can explain that evidence in plain language, but it cannot change a result, deploy code, approve a release, or contact users. The final decision always stays with the owner.

Challenges and lessons

The main challenge was avoiding vague AI judgments. We learned to separate deterministic evidence collection from AI explanation, so the safety checks remain reproducible and auditable.

The demo uses synthetic reservation-app fixtures only. It shows an unsafe build failing all three checks, an owner requesting repair, and the repaired build passing the same checks.

Outcome

IntentKeeper makes an AI-built app's intended behavior visible, testable, and under human control before release.

Built With

• boto3
• python
• strands-agents-sdk
• streamlit