Skip to content
Farvex.ai
Solutions Case studies Architecture Security Contact Pre-deployment checklist
Architecture

From a file landing to a booked transaction

Five layers, from the moment a file reaches us to the audit record written once it is booked. This page sets out what runs at each layer, how long it takes, and what happens on a bad day.

The route

The route every document takes

Each job passes through the same five layers in the same order. Nothing is skipped, and every layer records what it did.

  1. 01

    A file lands

    A PDF, a scan, an email or an API call reaches the gateway. Its signature is checked and it is queued before anything reads it.

  2. 02

    The page is read

    Vision models lift fields and tables off the page, handwriting, stamps and customs codes included.

  3. 03

    Your rules decide

    Checksums, VAT numbers and IBANs are verified against your business rules before a single value is written anywhere.

  4. 04

    The record is written

    It lands in Exact Online, SAP or your database as one committed transaction — all of it, or none of it.

  5. 05

    The run is logged and watched

    Throughput, latency and model drift reach your dashboards, and every step is appended to a write-once audit log.

10,000/s Requests the gateway takes
94ms Median time to read a page
99.98% Field-level accuracy
114ms p99 latency, end to end
Specifications

What runs where, and what a bad day looks like

The same five layers with the numbers we hold ourselves to, and the behaviour you get when something breaks.

Runtime, latency and failure behaviour for each of the five layers
Layer Runs on Typical time On failure
01 Ingestion gateway Managed edge nodes in EU regions 6 ms The payload stays queued and the sender gets a retry window.
02 Reading the page NVIDIA A100, INT8 TensorRT 84 ms The page goes to a review queue with the uncertain fields flagged.
03 Business rules Stateless workers, fixed state machine 12 ms The circuit breaker halts the job before anything is written.
04 Writing back Direct connectors and SQL bridges 210 ms The two-phase commit rolls back and the job replays from the buffer.
05 Monitoring and audit OpenTelemetry, Prometheus, Grafana Continuous An on-call engineer is paged inside the SLA in your contract.
In detail

The five layers, one at a time

What each layer runs on, and a sample of what it writes to the log as a single freight invoice passes through.

Layer 01

Ingestion gateway

Accepts PDFs, scanned TIFFs, IMAP mailboxes and REST calls at up to 10,000 requests a second. Every payload has its signature checked and is queued before a model sees it.

TLS 1.3 Token-bucket rate limiting Kafka event bus
Gateway log 202 Accepted
// EDGE INGESTION ROUTER
POST /api/v4/stream/ingest
Host: edge.farvex.ai
Authorization: Bearer fvx_live_994a...
Payload-Type: application/pdf
Event-Queue: KAFKA_BURST_BUFFER_01
Status: 202 ACCEPTED (6.2ms)
Layer 02

Reading the page

Transformer models tuned with TensorRT pull out tables, handwritten notes, customs HS codes and European IBANs at 99.98% validated accuracy. You never have to supply a training set.

TensorRT FP16 / INT8 Under 100 ms a page
Model run 84.2 ms
// DOCUMENT OCR EXECUTION
Runtime: NVIDIA A100 Tensor Core
Quantization: INT8 TensorRT
Batch-Size: 1 (Realtime Stream)
Inference-Duration: 84.2 ms
Checksum-Verify: ALL_PASSED
Layer 03

Business rules

Your rules run as a finite state machine: fixed transitions, no open-ended loops. When a check fails, a circuit breaker stops the job rather than letting the system guess.

Finite state machine Circuit breaker 3 s hard timeout
State transition PASS
// STATE MACHINE TRANSITION
State: RECONCILING_VAT_LEDGER
Condition: ChecksumMatch == True
Next-State: EXECUTE_EXACT_POST
Failsafe-Active: True (MaxTimeout: 3000ms)
Layer 04

Writing back to your systems

Two-way connectors for Exact Online, SAP S/4HANA, AFAS, Salesforce and SQL back ends. Tokens rotate by themselves and every write is a two-phase commit, so a half-finished posting cannot land.

Two-phase commit OAuth2 PKCE rotation Idempotent replay
ERP write 201 Created
// EXACT ONLINE REST BRIDGE
POST /api/v1/7100/journalentries
X-Correlation-ID: FX-SYNC-88214
Response: 201 Created
GL-Account: 7100 (Freight Costs)
Audit-Recorded: true
Layer 05

Monitoring and audit

OpenTelemetry pushes throughput, error rates and model latency to your dashboards. Every step is hashed into a write-once log you can hand an auditor as it stands.

OpenTelemetry Prometheus & Grafana SHA-256 audit log
Metrics Nominal
// OBSERVABILITY METRICS
Metric: latency_p99_ms = 114.2
Active-Agents: 32 worker nodes
Drift-Score: 0.002 (Nominal)
Hash: e3b0c44298fc1c149afbf4c8...
Deployment

Three places it can run

Three options. The software is identical in all three; only the ground under it changes.

EU managed cloud

We host and operate it in an EU region you pick. Your data stays in that region and we carry the on-call rota.

Your own cloud account

The same stack deployed into your AWS, Azure or GCP tenant. You hold the keys, the network rules and the bill.

On-premise

Runs on your own hardware, fully disconnected if that is what you need. Models ship as signed containers you install yourself.

That choice decides where the machines sit, not how the system behaves. The audit log, the monitoring and the fallback rules are identical in all three.

See how we handle security

Stack questions

What engineers ask us

No. Farvex runs beside your current systems and writes into them. Nothing has to be switched off to make room.

The job halts at the business-rules layer and goes to a review queue with the uncertain fields flagged. Nothing is guessed and written.

Yes. Every step is appended to a write-once log with a hash chain, and you get read access to your own traces and metrics.

Model and connector updates ship as signed containers and run against your own traffic on a staging pipeline before they go anywhere near production.

Exact Online, SAP S/4HANA, AFAS, Salesforce and HubSpot, plus direct SQL. Anything else is a connector we build during the project.

See what we ship

Want this walked through against your own systems?

A 45-minute technical briefing: we map one of your workflows onto these five layers and point out where it would break.