CaseGraph AI:
From Scattered Evidence
to a Traceable Case

Hey friends,

This week at Hacker House Goa 2026 we built CaseGraph AI, and I want to write down how it works and what building it taught me. It helps a person who was cheated or attacked online to report it clearly, and it helps an investigating officer see which earlier complaints connect to it.

The problem Scattered proof.

When someone is cheated online, the evidence is everywhere: chat screenshots, a payment receipt, a call log, and a story told out of order. Reporting usually means long forms and legal words, often while the person is scared and may have just lost money.

The officer on the other side has a different problem. Reading one complaint, they cannot easily see that the same phone number, UPI ID or website already appears in other complaints.

What it does Tell us what happened.

The person signs in with a name and mobile number and tells the story however they can: type it, speak it (Indian languages are detected automatically), or add screenshots, PDFs and documents.

As things arrive, AI reads them. Amazon Textract and Amazon Bedrock build a short summary, a dated timeline, key facts, and the phone numbers, UPI IDs, accounts and websites involved. Then it asks only for what is missing, one question at a time, and "I don't know" is always an acceptable answer.

At the end the person checks the story, downloads their own report (with their screenshots embedded and numbered) and submits, getting a reference number like CG-2026-000123. It is not only for scams: UPI and payment fraud, investment and job scams, loan-app harassment, sextortion, account hacking, identity theft, malware, fake profiles and more are all covered.

The case network Seen this before?

This is where TigerGraph comes in. On submit, the case and its identifiers go into a graph: phones, UPI IDs, bank accounts, websites, emails and handles become nodes, joined to every complaint that mentions them. A real query then returns earlier complaints that share something with the new one, with a score and the exact details they share.

The officer gets an inbox, a full case file with the timeline and evidence, a side-by-side comparison with any similar case, a small network view, and a printable dossier. For the demo we seeded about 70 fictional complaints across 15 crime types, including one bank account that shows up in investment, job and romance scams.

What building it taught me Do not let the AI guess.

1. Uploaded text is data, never instructions. People upload screenshots and documents, and any of it could say "ignore your rules". We wrap everything as data, and the AI's output must pass a strict schema before anything is saved.

2. Never show a number the person did not give. Any amount or ID that is not in their own material is dropped. Transaction IDs are copied exactly as printed, and an ID that a screenshot shows partly hidden stays partly hidden. It is never matched to anyone.

3. Split big AI jobs into small ones. One prompt did the summary, facts, timeline and questions. On a screenshot-heavy case it sometimes wrote a one-line timeline and forgot the call that started everything. Giving the timeline its own focused request fixed it.

4. Test the loops. Every answer re-ran the analysis, and every re-run produced new questions, so the person could never finish. Now questions only change when new material arrives, and a test fails if they ever get replaced mid-answer.

Honest limits What it is not.

CaseGraph is not an official police, government or banking service. It does not file an FIR, freeze accounts or recover money. Sign-in is name and mobile only in this version, so SMS verification and officer single sign-on are needed before real use. The search follows direct links only, not chains of complaints yet, and the AI can be wrong, which is why every item shows where it came from and the person can correct it.

Try it See it work.

The live demo is at casegraph-hhgoa.duckdns.org (officer view at /officer). The server is switched on for demos to save cost, so if it does not load it is switched off. The code is on GitHub. It is built with React, NestJS, PostgreSQL, Amazon Textract and Bedrock, Sarvam AI for voice, and TigerGraph for the case network.

If there is a problem close to home, build for it. The best hackathon demo is one that people understand in ten seconds.

Thanks to Hacker House Goa and TigerGraph for the push. 🚀