Ledger Copilot, in 90 seconds.
We're building the AI bookkeeper that watches everything — bank transactions, credit cards, books — and runs continuous audits + reports against a deterministic engine. Tested against six business types from a $850K coffee shop to an $85M housing authority. Modeled $138K annual savings on a $75M construction GC. $1.4B in projected addressable savings across ten Fortune-500 targets we've analyzed.
Dollar figures on this page are forward-looking projections and internal market-sizing estimates, not realized or guaranteed results.
Why now
For 100 years, the bookkeeper's value was throughput — txns per hour. AI changes that. The mechanical work (categorize, reconcile, post journal entries) is now machine work. The human role shifts to interpretation and strategy.
2024 IRS rule cut e-file threshold from 250 returns to 10. HUD penalty schedules are tightening. State agencies now actively pursue agencies with SEMAP scores below 60%. AI catches issues in-period, eliminating the year-end mystery.
Plaid alone gives us 12K institutions. Our stack adds OFX Direct Connect (no aggregator dependency), file upload (universal fallback), and an MX-ready adapter slot. We have 100% bank coverage with zero single-point-of-failure.
What's shipped
Not slides. Live, runnable, end-to-end. Click through any of these.
The architecture, briefly
Deterministic-first, LLM-second
All detection (anomalies, reconciliation, audit findings) is rule-based and provable. LLMs only narrate and translate — never make accounting decisions. This is the architectural pattern Rillet uses; it's the only one that's audit-defensible.
A multi-engine deterministic fleet (see /api/health for live counts)
Modular monorepo. Anomaly engine, reconciliation engine, audit engine, financial reports, file importers, bank connectors, activity log — each independently testable, each with explicit contracts.
Universal bank coverage
Plaid is primary. OFX Direct Connect (Schwab, Fidelity, USAA, Navy Fed) bypasses every aggregator. CSV/OFX/QFX file upload is the universal fallback. MX adapter slot ready for activation.
Provable continuity
Every audit produces a SHA-256 fingerprint over input IDs + finding codes. Re-run the audit later on the same data, get the same fingerprint. Downloaded backup JSON has its own content hash in the HTTP header. The financial picture is back-testable.
Get in touch
Investor inquiries, partnership conversations, or just curious — happy to share the deeper data.