AI that splits bank lines, asks when unsure, can't approve its own work, and logs every action on a tamper-proof audit trail.
Watch Jarvis break down a $3,912.44 PayPal transfer, match it to evidence, and hand it to the controller for approval.
Jarvis reads the bank line
Finds 4 accounts, balanced
Human required (403)
Hash-chained log
Propose splits with evidence. Challenge when something's unclear.
Review and approve. Jarvis can't post without you (HTTP 403).
Trace every number. Hash-chained audit log shows who, when, what.
Inspired by my boyfriend, an accountant. These are the real pain points he deals with.
Matching revenue and expenses to the correct period, even when the invoice arrives late or early.
Creating the financial package on time with accurate comparisons to prior periods.
Decoding cryptic bank descriptions to find the real vendor, account, and purpose.
Storing receipts, invoices, and approvals in a way that auditors can actually find them.
One company, many names: Claude vs Anthropic, Cursor vs Anysphere, same vendor, different entries.
Split, challenge, and simulate. Every number backed by a source.
Four layers of control, from entry to audit log.
When the AI tries to approve its own entry, the gateway returns "403 Forbidden".
SHA-256 audit log links each entry. Tamper any row and the chain verification shows exactly where it breaks.
Click any split to see the bank line, memo, amount, and attached receipt that support it.
Afterglow Labs is fictional so no client's books are exposed, but the code is real: single gateway, role checks, state machine, SHA-256 audit log.
Why this matters: The demo uses fictional data (Afterglow Labs, Sep 2026 scenario), but the controls are production-grade: single ledger gateway, HTTP 403 denial, state machine, and tamper-evident audit log.
In-browser simulation. No backend required.
Afterglow Labs (fictional AV touring studio), September 2026 scenario, Maya's Steam purchase, EUR crew deposit, Chase/PayPal/Ramp transaction data.
Single ledger gateway (only write path), role-based authorization (HTTP 403), state machine (new → proposed → needs_info → approved → posted), SHA-256 hash-chained audit log, challenge/defend/concede with precedent recording.
Afterglow Labs is synthetic so no real client's books are exposed, but it's modeled on how real books look:
What's real: The reasoning, the AI calls, the controls (403 denial), and the hash-chained audit log. This is not derived from any real company.
Common questions about Jarvis, the controls, and the demo data.
Every action appends to a SHA-256 hash-chained audit log with actor, timestamp, amounts, and state. Tamper any row and verification shows exactly where the chain breaks.
No. When Jarvis tries to approve its own work, the gateway returns "403 Forbidden". Try it in the Live Demo.
Simulated: Afterglow Labs (fictional AV touring company), September 2026 scenario, Maya's Steam purchase, EUR crew deposit.
Real code: Single gateway, HTTP 403, state machine, SHA-256 hash-chained audit log, challenge/defend system.
Why the synthetic data is realistic: Afterglow Labs is synthetic so no real client's books are exposed, but it's modeled on how real books look: QuickBooks-style chart of accounts (1010 Cash, 1250 Employee Receivable, 6410 Software, 6810 Bank & FX Fees), real bank-descriptor formats ("PAYPAL *INST XFER"), multi-source feeds (bank, PayPal, Ramp, GL), to-the-cent amounts including FX fees ($3,912.44 with $147.50 wire fees), materiality thresholds, written policies (P-0317), and the real close problems (bundled payments, personal spend on corporate cards, timing & accruals, duplicate vendors, vendor bank-change fraud).
What's real: The reasoning, the AI calls, the controls (403 denial), and the hash-chained audit log. This is not derived from any real company.
Built with Cursor (Claude Sonnet 4.5) and Grok Bot during Cal Hacks 2026. TypeScript + Three.js + GSAP for demos. Python backend in the repo.