Skip to Main Content

Accnomos

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Accnomos — Demo video

Accnomos is bookkeeping, VAT and household money in plain English, for UK small businesses, sole traders, landlords and households, built on Meta’s Muse models. Your books are files in your own OneDrive; we keep no database. The problem. A VAT-registered UK small business must keep digital records and file through Making Tax Digital software; someone has to turn a shoebox of receipts into those records. Accnomos asks whether AI can do the typing while the owner keeps control. The approach. AI proposes: snap a receipt on the Android app, upload it on the web, or let Outlook intake fetch it from email; a Meta model reads it into strict JSON. You approve: nothing posts until you tap Approve. Code calculates: arithmetic checks, ledger events, the nine VAT boxes, invoices, payments, credit notes and reports are deterministic, tested code. In Ask, every figure is checked against the books; an invented number is replaced by the facts. Your books, your OneDrive. The runtime is stateless: the ledger is append-only event files in your OneDrive; Books.xlsx is regenerated for your accountant. Disconnect and every file is still yours. What is built. Receipt capture, review and a second reading; contacts and ageing; invoicing with part payments and credit notes; bank statements, rules, reconciliation and an Open Banking feed; reports; audit trail; the VAT return with an HMRC sandbox flow; MTD Income Tax figures for sole traders and landlords; jobs, mileage, payroll; household budgets and goals; Ask, a hands-free assistant and chat actions that wait for Approve; web and Android apps. The Meta models. Muse Spark 1.3 on the Meta Model API answers first on the live server; Muse Glimmer 30B, Meta’s open-weights model on our own server, is the fallback and local mode. Muse Voice Transcribe hears spoken questions, SAM 3.1 crops receipt photos, Muse Image made the brand art, Muse Code helped write it. Measured, and its limits. On 200 synthetic documents, clean and messy, every field was right on 95 % and 96 % (Muse Spark 1.3), 85 % and 80 % (Glimmer); real photos score lower. Ask got 32 of 32 figures exact. Autopilot, off by default, made 17 false posts in 467 documents, all re-dated copies of posted receipts. Honest status. The Meta paths ran live on the public server on 4 Nov 2026, records committed. HMRC is sandbox only: the demo files to HMRC’s sandbox with a test user. The bank feed has met TrueLayer’s sandbox only; the Android app one real phone. Other features are mostly tested on fakes; demo data is fictional. Only the owner’s server predates the event; docs/HONESTY.md lists every gap. Try it with no sign-in: https://accnomos.com → Try the demo (guide: /judges). Android APK: GitHub Releases.

Team