Dachshunds — Auditable Forensic AI

Inspiration

Financial fraud is often treated as a classification problem: a transaction is marked as suspicious and assigned a score.

We wanted to build something different.

As engineering and physics students, we were inspired by the way complex physical systems are analyzed through flows, conservation, networks, accumulation, and changes in behavior.

In fluid dynamics, looking at a single point is rarely enough. You need to understand where something comes from, where it goes, how it moves through a system, where it accumulates, and whether it eventually returns.

We realized that financial activity can be investigated in a similar way.

A simplified financial balance can be represented as:

[ \sum \text{inflows} - \sum \text{outflows} = \Delta B ]

where (\Delta B) represents the change in the financial position of an account or entity.

We were also inspired by concepts such as concentration, dispersion, and entropy:

[ H = -\sum_i p_i \log(p_i) ]

We do not literally use thermodynamics to decide whether fraud occurred. Instead, ideas from fluid dynamics, conservation laws, graph theory, and information theory influenced how we designed the system.

We stopped looking at financial records as isolated rows and started asking:

Where did the money come from?
Where did it go?
Did it return to its origin?
Was a large operation divided into smaller ones?
Does the documentary trail match the monetary trail?
What evidence supports the hypothesis?
What evidence could prove us wrong?

That led us to build an auditable forensic investigation engine, rather than a simple fraud classifier.

What it does

Our system analyzes an unknown SQLite financial estate containing:

vendors
invoices
accounting ledger entries
bank transactions
purchase orders
contracts
employees
EFOS records

It investigates five fraud schemes:

Phantom Vendor
Kickback
Round Tripping
Threshold Splitting
Revenue Inflation

The system does not automatically accuse someone because a transaction looks unusual.

Instead, it follows a complete forensic process:

SQLite Estate
      ↓
Estate Validation
      ↓
Relationship Graph
      ↓
Money Flow Graph
      ↓
Fraud Detectors
      ↓
Signals
      ↓
Leads
      ↓
Investigator
      ↓
Evidence Registry
      ↓
Challenger
      ↓
Monetary Reconciliation
      ↓
Decision Engine
      ↓
Findings
      ↓
Auditable Case File