The AI drafts.
A person decides.
MuleTrace never files a report or closes a case on its own. Every AI suggestion is cited, every human decision is logged, and approved reports are locked, so you can show an examiner exactly what happened.
How MuleTrace keepspeople in control.
- Human approval on everything. Dismissing a case needs a written reason. Escalating, drafting and approving are analyst actions.
- The AI never files. Approval produces a PDF for the analyst to submit through the FIU’s own channel.
- Every sentence cited. Code removes any sentence whose sources are missing or not in the case evidence, and shows what it removed.
- Append-only audit log. AI suggestions, analyst decisions and STR edits are recorded with who and when.
- Locked, versioned reports. Approved STRs cannot be edited; a change creates a new version.
- Fails safe. If the AI fails, cases fall back to a rules-only priority labelled as such, and earlier drafts are kept.
- Scoped AI tools. The case assistant can only read accounts linked to the open case.
- Data is never instructions. Customer text is passed in a marked data block and cannot change what the AI is told to do.
Where doesthe data live?
The prototype runs on Google Cloud in the Mumbai region (asia-south1): Cloud Run for the API and overnight triage job, Firestore for decisions and the audit log, and Cloud Storage for approved report PDFs. It uses synthetic data only.
For production, MuleTrace is designed to run the same prompts on Gemma, Google’s open model, inside the institution’s own servers or in-country cloud. That keeps customer data in Sri Lanka, in line with the Personal Data Protection Act, 2022.
| Need | Prototype |
|---|---|
| Language model | Gemini 2.5 Flash on Vertex AI |
| API and overnight job | Cloud Run and Cloud Run Jobs, asia-south1 |
| Decisions, drafts, audit log | Firestore, asia-south1 |
| Approved STR PDFs | Cloud Storage |
| Sign-in | Firebase Authentication + allowlist |
| Production plan | Gemma in-country or on-premises |
What MuleTraceis not, yet.
Security:your questions.
Last updated 13 October 2026
Where is data stored?
The prototype runs on Google Cloud in the Mumbai region (asia-south1), with synthetic data only. For production, MuleTrace is designed so the model can run on Gemma inside the institution’s own environment or in-country cloud, to meet data-residency needs under the Personal Data Protection Act, 2022.
Who can sign in?
Analysts sign in with Google through Firebase Authentication, and the API also checks each email against an allowlist. Requests without a valid token get a 401 response; signed-in users who are not on the list get a 403.
Can an approved STR be changed?
No. An approved report is locked. Any later change creates a new version, and both the edit and the approval are recorded in the audit log.
Does MuleTrace hold certifications such as ISO 27001?
Not yet. MuleTrace is a prototype and makes no certification claims. Security requirements for each pilot are agreed with the institution in advance.
Which model does MuleTrace use?
Gemini 2.5 Flash on Vertex AI in the prototype. The model sits behind a single interface, so switching to Gemma for an in-country deployment is a configuration change rather than a rewrite.
Run a piloton your own alerts.
We are looking for a small number of Sri Lankan institutions to try MuleTrace on their own historical alerts. Tell us about your team and we will set up a walkthrough.