Inspiration
Money moves through many hops: wallet, processor, bank rail. When something goes wrong, nobody can prove what happened. Disputes become "your word against ours," and reconciliation is done by hand, after the damage is done.
Now AI agents are starting to spend money on our behalf, and the trust gap is about to explode. How do you prove an agent stayed within what you approved?
We wanted a ledger where every record can be verified by anyone, and where AI explains problems before they turn into disputes.
What it does
LedgerTrue is a hash-chained, AI-assisted ledger built on the PayPal sandbox. Trust every transaction. Verify every record.
- Tamper-evident ledger: every record is cryptographically linked to the previous one, so any edit visibly breaks the chain.
- Real PayPal sandbox payments: live payment flows are written into the ledger as signed receipts.
- Proof of Intent: records what the payer meant to do, then compares it with what PayPal actually did. A QR-verifiable Proof Card lets anyone check the result.
- PayBlackBox: a signed receipt chain, a hop-by-hop payment tracker, rail health warnings, and a one-click dispute claim pack.
- Ghost Balance: a reconciliation engine that detects balances that exist on one side but not the other.
- AI layer (NVIDIA NIM): explains anomalies in plain language and drafts dispute evidence.
Our core design rule: the AI never decides what is true. Hashes, signatures and deterministic rules decide. The AI only proposes, flags and explains from the ledger's evidence, so it can never hallucinate a proof.
How we built it
- PayPal developer platform (sandbox): executes real payment flows whose results become signed ledger records.
- AI model via NVIDIA NIM: anomaly explanation and claim-pack drafting, grounded strictly in ledger evidence.
- Proof engine: hash chaining and signatures, with shared crypto and rules modules so the browser and server verify records identically.
- Node 20+ server with a zero-framework frontend in plain HTML, CSS and JavaScript.
- Quality: a test suite, documentation, and an open-source licensed public repo.
Challenges we ran into
- One definition of "valid": keeping client and server verification identical, so a valid chain means the same thing everywhere.
- Honest modeling: representing payment hops truthfully when the sandbox only simulates some of them.
- Grounded AI: forcing the model to explain from evidence instead of guessing.
- Scope: fitting a full, demonstrable system into hackathon time.
Accomplishments that we're proud of
- A ledger you can tamper with live and watch it flag the break.
- A dispute claim pack generated in one click from signed evidence.
- Shared crypto and rules code with tests, so verification is reproducible.
- A real PayPal sandbox flow, not mocked screens.
What we learned
- Trust comes from verifiability, not from asking users to believe a dashboard.
- AI is most powerful as an explainer on top of evidence, never as the source of truth.
- Small, boring invariants (hashes, signatures, tests) beat flashy features.
What's next for LedgerTrue
- Live PayPal production rails and multi-party ledgers (merchant, processor, bank).
- Anchoring chain checkpoints to a public timestamping service.
- Merchant dashboards for dispute teams, and an open verification API.
- Deeper Ghost Balance detection with learned anomaly models.
