Problem
A payout can be wrong even when every individual record looks plausible.
Finance operators need to compare expected and received amounts, isolate mismatches, and retain evidence for every decision.
Selected project · LedgerGuard AI
Finance operations that can prove where the numbers diverged.
LedgerGuard reconciles payouts with deterministic rules, isolates exceptions, uses bounded AI to organize the evidence, and keeps the final decision with a human reviewer.
Core principleDeterministic financial logic
AI boundaryEvidence-bounded exception analysis
ControlHuman-reviewed consequential actions
Solution
Separate the calculation from the explanation.
LedgerGuard keeps reconciliation deterministic, asks AI to interpret only bounded exception evidence, and gives the reviewer the final word.
Complete investigation
Why did £120 become £76.50?
The 82-second demo follows the mismatch, deterministic calculation, AI hypothesis, human decision, and audit trail.
Synthetic demonstration — no production financial system or customer data is connected.
Architecture
Financial truth first. Interpretation second.
Each stage narrows the problem while preserving the evidence required by the next one.
- 01Financial data intake
Orders, refunds, fees, and payouts enter one reviewable operating view.
- 02Deterministic reconciliation
Rules calculate expected values and match routine transactions without an AI model.
- 03Exception isolation
Only mismatches move forward for investigation and human attention.
- 04Bounded AI analysis
Claude organizes the evidence already attached to an exception and proposes a hypothesis.
- 05Human decision
A reviewer approves or rejects consequential actions; money is never moved automatically.
- 06Audit trail
Every run, attempt, status, and decision remains traceable.
Key engineering challenges
Use AI without making the ledger probabilistic.
Keep financial truth deterministic
Expected payouts are calculated by rules, not inferred by a language model.
Bound the AI context
Analysis is limited to evidence already associated with the exception under review.
Protect data access
Server-side credentials and row-level security keep privileged data out of the browser.
Prevent duplicate work
Workflow state and retry-aware execution reduce repeated operational actions.
Make every decision reviewable
Correlation, timing, outcome, and reviewer activity form one evidence trail.
Demonstrate safely
The public product uses synthetic data and does not represent a connected financial system.
Next.jsTypeScriptSupabasePostgreSQLRLSEdge Functionsn8nClaudeVitestVercel
Portfolio-safe proof
A working product demo, with its boundaries stated clearly.
The public demo shows reconciliation, exception analysis, human review, and traceable execution using synthetic records.
Illustrative figures demonstrate product behavior; they are not production financial results or customer metrics.
Explore the project