A fraud score should start an investigation, not end one. In CaseVera, a challenge project by Shaurya Upadhyay, a statistical model signals suspicion while graph relationships, retrieved context, evidence checks and policy rules shape what action—if any—is justified. The distinction matters: a high score is not proof, and an uncertain case can call for verification rather than a fraud verdict.
What does it mean for an AI system to investigate fraud?
CaseVera was built for the TigerGraph × Hacker House Goa 2026 challenge. Upadhyay describes it as an evidence-first fraud investigation agent: it gathers and evaluates information around a flagged case, then recommends a next step. It does not treat the model’s suspicion score as the decision itself. The project account was published September 24, 2026, and its examples and reported metrics are the author’s claims, not independently audited results. Read Upadhyay’s CaseVera account.
As an Amazon Associate I earn from qualifying purchases.
The design separates four things that are easy to conflate: suspicion, evidence, policy and approval. A model can raise suspicion; it cannot, by that fact alone, establish what happened, determine what policy permits or approve a high-impact action.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteHow does CaseVera move from a signal to a recommendation?
Upadhyay describes the sequence as: trigger, investigate the graph, retrieve context, assess evidence, handle uncertainty, request evidence if needed, apply policy, recommend the next best action, stop and persist case memory. Each part has a bounded role:
#1 Best Overall
- Statistical model: contributes a suspicion prior, not a final finding.
- TigerGraph and GSQL: expose relationships and facts through graph queries. TigerGraph MCP serves as the transport for graph tools.
- GraphRAG: retrieves historical cases and policy or regulatory context. A similar old case can inform an investigation, but cannot prove what happened in a new one.
- Evidence engine: assesses which current facts support or contradict a conclusion and whether more evidence is needed.
- Deterministic policy rules: map findings to permitted actions and approval routes.
- Case memory: persists investigation information so cases can be reconciled and revisited.
This division is the project’s central safeguard: graph facts and retrieved context inform the evidence assessment, while policy—not a generated answer—constrains the action. The author’s principle is concise: “Generation should improve the interface. It should never manufacture the evidence.”
What does contradictory evidence change?
HHG-007: high suspicion, but no supporting fraud signals
In the author’s benchmark example HHG-007, the model expressed about 91.8% suspicion. The case reportedly had zero fraud-supporting signals and five legitimacy signals. Its transaction amount, region and product behavior were described as normal, and the customer had hundreds of similar historical charges. CaseVera therefore remained uncertain and recommended “Verify first,” rather than treating the score as a finding. In a simulated denial scenario, it would lead to a conditional block.
Rank #2
That example shows why a score needs interpretation: 91.8% is the project’s reported suspicion figure, not an independently established probability that the customer committed fraud. The legitimacy signals do not prove innocence either; they are evidence to weigh and a reason not to jump from model output to an irreversible action.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →HHG-014: relationships can expose a wider pattern
For HHG-014, Upadhyay says a device-fingerprint traversal found one fingerprint connected to 24 customers and four connected cards. A transaction row viewed alone may not reveal those links. A graph can make connections across entities visible, giving an investigator a pattern to examine. But the connection itself still needs interpretation: shared infrastructure or a repeated relationship can be a lead, not automatic proof of fraud.
That distinction also applies to historical-case retrieval. Upadhyay puts it this way: “Precedent is context. It is not proof.”
Who decides whether to block, verify or escalate?
CaseVera’s policy layer is described as routing high-impact actions through approval levels labeled auto, L1 or L2. The agent may recommend a restricted action without representing that a human has approved or carried it out. This separates a recommendation from an authorization and an authorization from an executed action.
The same boundary applies to suspicious activity reporting. The project says SAR generation is controlled by policy rather than triggered by an LLM. FinCEN’s SAR FAQ is an official reference for SAR questions; neither that FAQ nor CaseVera’s account establishes that this challenge project satisfies regulatory requirements or is compliant for deployment.
Free tools Windows power users keep installed
One-click scans. No signup required.
There is a practical reason to keep review and evidence in the loop: flagged activity can create investigative work, including false positives. TigerGraph’s own AML article discusses further human review of flagged activity and the burden of false positives; it is vendor context, not independent validation of CaseVera or a neutral comparison of approaches. TigerGraph’s AML investigation discussion.
Best Value
![Free Fling File Transfer Software for Windows [PC Download]](https://m.media-amazon.com/images/I/41Vq6ZqHfjL._SL160_.jpg)
- Intuitive interface of a conventional FTP client
- Easy and Reliable FTP Site Maintenance.
- FTP Automation and Synchronization
What did the project validate—and what did it not?
Upadhyay reports the following engineering checks for the challenge project. They describe output handling, persistence and tool behavior, not whether the system correctly identifies fraud:
| Reported check | What the author reported | What it establishes |
|---|---|---|
| Answer files | 20/20 produced and schema-validated (Upadhyay, 2026) | Files met the project’s schema checks; this is not a fraud-accuracy measure. |
| Persisted cases | 20/20 persisted and independently reconciled (Upadhyay, 2026) | Stored case records were reconciled; this does not establish correct case outcomes. |
| Evidence vertices | 273 Evidence vertices (Upadhyay, 2026) | A reported graph-content count, not a quality or accuracy score. |
| Graph-query parity | 7/7 queries returned identical normalized results over REST and MCP (Upadhyay, 2026) | The two paths returned matching normalized results for those tested queries. |
| MCP runtime | 96 runtime calls and zero MCP failures (Upadhyay, 2026) | No MCP failures were reported in those calls; broader reliability is not established. |
| Semantic comparisons | Zero semantic differences between MCP and REST, and zero between MCP and the approved release (Upadhyay, 2026) | The reported comparisons matched semantically; they do not validate the underlying fraud judgments. |
The author says benchmark labels were hidden and does not claim an accuracy number. Without an available independent evaluation against known outcomes, these counts cannot show predictive accuracy, false-positive rates or real-world fraud outcomes. A correctly transported or persisted answer can still be wrong.
Is CaseVera a production-ready fraud system?
The account describes a challenge project and an architectural foundation, not an established production deployment. Its customer replies were simulated, its browser snapshot was deliberately read-only, and the author identifies real analyst and customer interaction, calibration, broader fraud typologies, streaming triggers, analyst feedback and retrieval of newly written case memory as future work. Those boundaries matter when interpreting the examples and validation counts: they show how the design was exercised in the project, not how it performs in live operations.
The project’s strongest contribution is therefore a decision-making boundary, not a demonstrated accuracy result. A score can trigger scrutiny; connected data can supply leads; contradictory facts can change the next step; policy can limit what is allowed; and a human approval route can remain explicit. As Upadhyay writes, “The most intelligent system is not always the one that answers fastest. Sometimes it is the one that knows when it does not know enough.”
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
