A more cinematic version of the original demo page. Same core story, but presented like a live operating system: invoice intake, AI reasoning, Business Central posting, bank reconciliation, and continuous feedback loops.
This version keeps the original content but reframes it with a stronger presentation: more motion, more hierarchy, and visuals that feel like a real autonomous finance control room.
Email arrives, PDF is read, lines are classified, the journal is posted to Business Central, the document is attached, and a notification is sent. The full loop finishes in seconds.
A six-tier matching engine starts with deterministic rules and only escalates to reasoning models for the difficult edge cases, including partials, split payments, and intermediaries.
Hours and expenses become structured work reports, then clean PDF invoices, then posted sales invoices in BC, then outbound delivery to the customer without leaving the workflow.
Open vendor payables are converted into payment batches, enriched with learned bank account details, and submitted directly through banking integrations.
Unknown vendors are checked against VIES, official registration data is retrieved, and the master record is created with the right posting groups and identity fields.
High-confidence invoices auto-post. Lower-confidence or mixed-category cases route into review. Human decisions become training signals for the next pass.
The left side animates the operational sequence. The right side shows the signals the system uses to decide whether to post, enrich, or escalate.
Local models cover cheap, repetitive matching. Cloud models handle perception and reasoning. The orchestration layer chooses when to spend money and when not to.
The system is designed so AI spend only appears when deterministic logic runs out. That changes the economics completely.
About ten hours per month, office-hours throughput, repetitive work, and the same edge cases resolved over and over again by hand.
Fine for clean inputs, weak on ambiguous references, partial payments, exceptions, and cross-document context that needs judgment.
Most cases stay free and local. AI only handles the tail. The result is a 24/7 process that gets sharper as it observes corrections.
Transactions enter the cascade and exit as soon as a sufficient match is found. Most entries never touch an external model.
| Tier | Engine | Method | Coverage | Speed | Cost |
|---|---|---|---|---|---|
| 1 | Deterministic | IBAN + exact amount, invoice reference match. | ~50% | <1ms | Free |
| 2 | Fuzzy 6-signal | IBAN, reference, embedding, name, amount and date combined. | ~25% | ~3s | Free |
| 3 | Split and partial | Combined payments, part-payments and credit note offsets. | ~8% | ~100ms | Free |
| 4 | Exceptions | Rounding, bank charges, discounts and tolerated mismatch logic. | ~5% | ~50ms | Free |
| 5 | AI panel | Parallel DeepSeek and Grok reasoning with consensus. | ~10% | ~12s | €0.001 |
| 6 | Journal generation | GPT-4o-mini generates BC journal shape and GL context. | Always | ~2s | €0.0005 |
Keep this page as the cinematic overview and link out to more focused demos for specific flows.
Worker service and Blazor UI.
v2.0 and OData v4 APIs.
Email monitoring and triggers.
EF Core and local state.
Local embeddings and retrieval.
Encrypted secret handling.