Your fraud analyst sees rows. Fraud is a shape.

How we built CaseFile, a graph-walking investigation agent, in one hackathon day — and what it taught us about context for AI agents.

---

An alert lands on your fraud team's queue: $8,150.50, account FR-04 → FR-02, both customers are new. Your analyst opens FR-04's file. Name, country, account age, recent activity. Ordinary.

Then they open FR-02's file. Also ordinary.

The products told them all of this in two separate windows — and both windows lied by omission, because the fraud isn't in either record. It lives between them:

Five other "customers" — all weeks old — log into their accounts from the same two phones as FR-04. There is no customer named on those accounts that you ever met at a branch. One innocent woman shares a device with them accidentally, and she has ten months of boring grocery payments. And the money, once it lands, walks FR-04 → FR-02 → FR-05 → FR-03 and gets paid out to the same electronics merchant.

That is a shape. Rows don't have shapes — graphs do.

The agent that walks it

CaseFile is our hackathon agent (Graph Hacks by WeMakeDevs × FalkorDB). Its LLM has no memory of any customer in its context window. What it has instead is nine tools, each one a Cypher traversal over FalkorDB:

Tool	What it hands the agent
get_alert / get_open_alerts	the paperwork: who, what, why flagged
account_profile / account_connections	one hop around a suspect
shared_endpoints	other accounts behind the same device / IP / card
trace_money_path	follow transfers N hops
shortest_path	cheapest evidence trail between two accounts
ring_extent	how big is the cluster around this account
close_alert	the agent's write action — verdict into the graph

We deliberately gave the model no customer data in the prompt. It investigates by choosing queries and reading back results, so every fact in its final verdict is reproducible: show the tool call, show the Cypher, you have shown the reasoning.

The point where the demo gets applause is the decoy. The graph that incriminates six accounts also tabs one completely innocent woman as a device sharer. A keyword/ML-risk-score system would flag her and move on, generating a false positive (and, honestly, an unfair suspension). The agent instead runs her profile next to the other five, sees ten months of normal behavior vs. five new flag-plagued accounts, and explicitly says: "ruled out — ordinary history, plausible shared family device." That's reasoning over relationships and their evidence, not pattern-matching on a connection existing.

Why the graph is the context layer, not a sidecar

Every serious agent question here is about connections: which accounts are linked to this transaction, what did the money do next, who else was holding this device. Our agent loops entirely on:

MATCH (a:Account {id:$id})-[:USES_DEVICE]->(d:Device)<-[:USES_DEVICE]-(b:Account)
RETURN d.id, collect(DISTINCT b.id)

Three hops of that — and the ring is simply visible. And because FalkorDB keeps the answer to close_alert in the graph along with the fact pattern, the next session starts from an evidence-labeled graph, not a blank slate. The database is the agent's memory and its working context.

The verdict shouldn't be a paragraph of LLM prose you have to take on faith. Ours is a trail: 7 tool calls, each one reproducible, ~10 ms of graph time total. Agents are only as good as the context they work with — ours works inside the shape of the problem.

---

Built with FalkorDB + Python in one day for Graph Hacks. Full code: github.com/sksaddam007/casefile. To reproduce: CASEFILE_MOCK=1 ./demo.sh.