Archon — AI Financial Intelligence for SMBs

Drop a month of bank, payroll and invoice documents — Archon's AI reads them and a deterministic engine reconciles them into an auditable P&L close for SMBs. On Vercel + AWS.

Comment

Inspiration

Every month a small business closes its books, and the truth of that month is scattered across documents that don't talk to each other: a bank statement, a payroll register, individual payslips, supplier invoices, receipts. Each tells only part of the story, so owners (and their accountants) correlate them by hand, every month.

The insight that started Archon: a single payroll event produces three documents that disagree on purpose. The bank confirmation shows the net that left the account (€3,994.74). The payroll register holds the true employer cost, gross pay plus employer social-security (€6,930). That €2,935 gap, in our demo it's Greece's social-security contributions (IKA, the national social-security fund) plus withheld income tax, but every country has its own version of this employer-cost wedge, is invisible until you correlate the register with the bank transfer. Bank-only accounting simply can't see it. The correlation problem is universal; only the local rules change. We wanted software that does that correlation for any SMB and tells you one thing: are your books complete and reconciled, or are you missing a document?

What it does

Archon ingests a month's raw business documents and produces a boardroom-ready P&L close.

  • Drop a document on the dashboard and watch eight agents work in sequence: the Extractor reads it live with AWS Bedrock vision, then the engine classifies it, links related documents into one financial event, validates them against each other, and recomputes the close, the affected tiles flash and update in front of you.
  • Verification-gating: four cross-document checks (R1–R4) must pass before a close is trusted; a built-in stress-test deliberately corrupts a field to show the completeness check catch a missing document.
  • Documents-first search: find any invoice by number and date (e.g. AR-HA-003-001 · Sales invoice · 2026-01-22 · €3,304 · paid), not just rolled-up totals.
  • A self-guided tour walks any visitor through the whole flow.

Who pays (B2B SaaS). SMB owners and the accounting firms that close their books, a per-entity monthly subscription, with a multi-client tier for firms that close dozens of SMBs a month. The ROI is concrete: Archon turns a full day of manual document-correlation per monthly close into minutes, an accountant-day saved every month, per entity, and it surfaces costs (like the employer-cost wedge) that bank-only bookkeeping silently misses.

Live: https://h0-archon.vercel.app · Code (MIT): https://github.com/upgradedev/h0-archon

How we built it

The "zero stack": Next.js 16 / React 19 on Vercel for the front and API (Vercel Functions), with the real work on AWS:

  • Amazon DynamoDB, single-table design (REPORT + ACTIVITY items) is the source of truth for every close and product event. We chose DynamoDB over Aurora because the access patterns are simple and known (latest close, closes over time, recent activity), they map cleanly to partition + sort-key queries with single-digit-ms reads, no connection pool to manage, and per-request pricing that's near-free at this scale; a serverless front (Vercel Functions) pairs naturally with a serverless, connectionless datastore.
  • Amazon Bedrock (Claude Sonnet 4.6), vision extraction via Converse document blocks (PDF-direct), ~96.7% measured field accuracy on a labelled corpus.
  • Amazon OpenSearch, a CQRS read-model fed from DynamoDB that powers documents-first search; it never computes a canonical number.
  • The guiding principle: the AI reads, a deterministic engine computes. No LLM ever touches a number, the same inputs always produce the same books, and every figure traces back to a source document. That's what makes the close auditable.
  • Terraform IaC for the whole data tier; GitHub Actions CI runs gitleaks + CodeQL + typecheck/test/build + a live smoke that hard-asserts db_mode=aws-dynamodb.

Challenges we ran into

  • Drawing the AI/determinism line. Letting an LLM "just compute the totals" is tempting and wrong, it kills auditability. Confining Bedrock strictly to reading and routing every number through deterministic rules was the key design discipline.
  • Single-table modeling. Getting a collision-proof sort key (<generated_at>#<event_id>) and the access patterns right took iteration, but removed an entire relational layer.
  • Deploy ops. Vercel's Git auto-deploy stalled mid-sprint; we built a manual GitHub Actions deploy workflow (Vercel CLI + secrets) as a reliable fallback.
  • Trust, verified. We caught the production app briefly serving stale data during a video render by checking the output frame-by-frame, a reminder that "it deployed" isn't "it's correct."

Accomplishments that we're proud of

  • A live, working product on the zero stack, not a mockup.
  • 96.7% measured extraction accuracy, with every downstream number deterministic and auditable.
  • The payroll-truth insight made tangible: €3,995 vs €6,930, with the €2,935 wedge surfaced only once documents are correlated.
  • Real engineering rigor: documents-first search with number+date, Terraform IaC, and a CI pipeline that gates on live AWS behavior.

What we learned

  • Let the LLM read, never compute, it's the cleanest way to get AI leverage without losing auditability.
  • CQRS keeps search rich (OpenSearch) while the source of truth (DynamoDB) stays untouched.
  • A serverless front pairs naturally with serverless data, Vercel Functions + DynamoDB, no pool, no idle cost.
  • Verify the artifact, not the step, frame-checking the demo caught a real data bug a green deploy hid.

What's next for Archon : AI Financial Intelligence

  • More document types (VAT returns, contracts, receipts) and direct bank-feed ingestion.
  • Accountant workflows, review queues, approvals, period locking.
  • Analytics scale-out on the existing OpenSearch read-model (trends, anomaly detection).
  • Global from day one: multi-entity / multi-currency and pluggable country-specific tax packs for SMBs worldwide, starting with Greece, its IKA social-security rules and MyDATA (the country's mandatory e-invoicing / e-bookkeeping platform), as the proof market, since the correlation engine is the same everywhere and only the local rules differ.

Built With

Share this project:

Updates