GreenLedger — AI bookkeeping engine (full IP)

Gallery

About

1) WHAT IT ISGreenLedger is a tested prototype of a deterministic double-entry bookkeeping engine built on the principle: AI proposes, humans decide, the database enforces. An AI categorizer proposes an account for each transaction with a confidence score; a deterministic policy engine routes each proposal to auto-post or to a human review lane based on confidence, amount, cumulative caps, sensitivity, and a per-account error rate; and PostgreSQL — not application code — enforces the accounting invariants (balanced entries, append-only journals, mandatory provenance, period locking, and a tamper-evident hash chain). Every number carries full provenance: one endpoint returns the complete paper trail for any line (line to entry to routing decision to proposal to source document to reviewer to hash), rendered verbatim in the UI.2) WHAT'S INCLUDEDFull outright IP transfer of a single-buyer sale. You receive: the complete Git repository at branch release/ip-sale including full commit history from first commit to sale; all application source (FastAPI backend with policy engine, categorizer, ledger and close services, and LLM client; Vite + React frontend with design tokens and API client; ordered PostgreSQL schema and bootstrap; root prototypes that form the behavioral contract; scripts and Docker/CI infrastructure); the binding specification and design record (spec, documented deviations, deferred-conformance register, and a phase-by-phase README build narrative); the full pytest suite of 57 passing tests plus 1 skipped (the skip is a live-LLM smoke test), spread across 11 modules, with a GitHub Actions CI gate (Postgres 16 service container to schema reset to pytest to vite build); design tokens and vendored IBM Plex fonts; synthetic demo/seed data for a fictional company (no real business's books); and buyer documentation — the asset schedule, this buyer-diligence briefing, and an IP-assignment draft.3) TECH STACKDatabase: PostgreSQL 16, with schema-enforced invariants (deferrable constraint triggers, a SECURITY DEFINER hash-chain trigger, and role GRANTs). Backend: Python 3.11, FastAPI, Uvicorn, psycopg 3 with connection pooling, Pydantic 2. AI categorizer: Anthropic API (Claude Haiku), temperature 0, strict-JSON output, confidence capped at 0.90, with a pattern-memory layer in front of the LLM. Frontend: Vite + React 18, Recharts, and vendored IBM Plex fonts (no external fetch). Tests/CI: pytest (57 pass / 1 skip) on GitHub Actions with a Postgres 16 service container. Local infra: Docker Compose (Postgres + API) plus a no-Docker local Postgres runner. Runs end-to-end locally in about ten minutes.4) HONEST LIMITATIONSThis is a prototype, not a production SaaS. Known limitations (from the buyer-diligence doc, verbatim):Auth is a header, not a system. The acting role is the X-GreenLedger-Role request header. There is no authentication, no user identity, no session, no SSO, no authorization beyond role-lane checks. A real deployment must add an identity layer in front.Demo data only. No real ingestion. The seed is synthetic.No bank feeds / no external financial integrations. No Plaid, no QBO/Xero, no processor settlement un-netting. Ingestion is modeled close to one-line-per-document.The live-LLM path needs the Anthropic SDK (now pinned) and an API key. The tests mock the client, so the gate needs no network or key; running the real categorizer additionally requires an ANTHROPIC_API_KEY.No SOC 2, no security audit, no compliance certification of any kind.Legal / CPA-of-record / compliance structure is research only. The spec describes an accountability model; none of it is implemented or operational. This is not accounting or legal advice and must not be represented as a compliant bookkeeping service without the appropriate professional and regulatory structure.Single-node assumptions. Advisory-lock hash chaining and the pooling model are designed for a single Postgres primary; horizontal scale is unproven.5) SALE TERMSOne-time outright IP transfer to a single buyer, completed via escrow. The seller retains portfolio-reference rights — the right to reference, describe, and display non-confidential aspects of the work (including screenshots and architecture) for portfolio and professional purposes after transfer; this carve-out does not include the right to resell, relicense, or operate the assets. The code repository stays private until the sale closes. No domain, trademark, customers, revenue, patents, or production infrastructure are included (see the asset schedule for the full what-transfers / what-does-not list). Development provenance is disclosed openly: GreenLedger was built AI-assisted using Claude Code under human-gated phases, and the full commit history transfers intact.