From separate reports to one case graph, in five steps
A schematic of the method. The five reports and the links between them are made up; the live demo itself runs on synthetic reports, not real ones.
A lead is a connection across cases, with a citation behind it: the reports it comes from.
- A ringed dot on a card is a person named in that report. A dashed line ties names found to be the same person.
- Red string is a link in the case graph: one person, strung to every report that names them. The lead is the heavier string, and the reports it cites are outlined.
The problem and the idea
Links police first information reports (FIRs) into a searchable case graph, so investigators can see connections across cases with a citation behind every lead.
What I built
- LLM entity extraction with spaCy and regex fallbacks
- Fuzzy person resolution across reports
- Neo4j case graph plus Qdrant semantic search
- FastAPI backend with about 20 endpoints and NCRB hotspot data covering 783 districts
- Demo runs on synthetic FIRs
The result
- 783Districts of hotspot data
- ~20API endpoints
From a report to a lead
A made-up example. These two reports, their names and places are invented to show the steps; the live demo runs on synthetic reports, not real ones.
The connection comes out as a lead, and every lead carries its citations: the lines in the reports it rests on.
FIR 1
The complainant states that on the night of the theft a man named Ravi Kumarperson was seen near the market at Old Townplace with a second man, Sureshperson.
FIR 2
A shop owner reports a break-in. Witnesses name R. Kumarperson · same as Ravi Kumar, who was seen earlier with Imranperson near the bus depotplace.
- Suresh is named in FIR 1.
- Ravi Kumar is named in FIR 1 and FIR 2.
- Imran is named in FIR 2.
LeadFIR 1 and FIR 2 share one person: Ravi Kumar, written “R. Kumar” in FIR 2. Suresh and Imran are each one step from him.
- FIR 1
a man named Ravi Kumar was seen near the market
- FIR 2
Witnesses name R. Kumar
Stack
- Python
- FastAPI
- LLM
- spaCy
- Neo4j
- Qdrant