Back to Spectrum Labs Project Detail
Open B00KS
An open-source, self-hostable bookkeeping app that turns a receipt into a balanced journal entry — demonstrating guarded AI extraction, financial data integrity, and production operations for sensitive data.
Positioning
Applied AI where mistakes cost money: every model output is constrained, verified, and reviewed before it touches the ledger.
Overview
What this project demonstrates
Open B00KS is a bookkeeping tool for small operators who want their books kept as expenses happen, without SaaS rent or ERP overhead. Upload a receipt and it suggests the vendor, account, and payment; a person confirms, and it posts a balanced journal entry. AI is optional and narrowly scoped: each model call does one task against a strict schema, Go code checks the result against bookkeeping invariants, and low-confidence results wait for review. It is released under the AGPL and runs in production for Spectrum Labs’ own books.
Core framing Books recorded the moment an expense happens
Good Fit For
Finance and back-office tools handling sensitive data
Document extraction pipelines that must be accurate, not just plausible
Products that need AI suggestions with human approval built in
Self-hosted or privacy-sensitive deployments
Core Delivery Areas
Double-entry accounting with balanced-entry enforcement and statement reconciliation
Bank and card CSV import with duplicate detection
Postgres-backed job scheduling with leases, safe across replicas, with no external queue
Automated scheduled database backups to object storage with retention
Platform Stack
Go API, worker, and scheduler services on Postgres
SvelteKit and Svelte 5 single-page app served by nginx, with the API on the same origin
Cloudflare R2 for receipt storage and Typesense search with an in-database fallback
Helm on Kubernetes, OpenBao for secrets, OpenTelemetry, Prometheus, and Grafana
AI Pipeline
Decomposed pipeline: transcribe, extract, match vendor, classify account, build entry, then gate on confidence
Model chooses only from known accounts and vendors, never free text, and results are verified in code
Vendor learning means one AI call per new vendor ever, plus batch mode for lower-cost bulk processing
Applied Product Capabilities
Receipt capture with text-layer PDF reading before any model call
Profit and loss, balance sheet, and general ledger reports
Tax-pack export with readiness checks for handoff to a preparer
Business chart-of-accounts templates for common small-business types
Mileage tracking and home-office allocation
Benchmark harness that measures pipeline accuracy against hand-verified data
What A Client Could Hire Spectrum Labs To Do
Add AI document extraction to a workflow without letting the model make final decisions
Build finance or compliance tools with integrity checks enforced in code
Ship a self-hostable, open-source product with a clean deployment story
Run sensitive internal tools on private infrastructure with zero public API surface
What’s Next
Bank feed imports, multi-currency support, and QuickBooks and Xero exports
A mobile app for capturing receipts on the go