Solutions — six systems we build, wire in and keep running

At a glance

The six, side by side

Speed figures are medians measured end to end on production traffic.

What each of the six modules takes in, what it returns and how fast it runs
Module	Takes in	Returns	Typical speed
01 Workflow automation	Events from mail, CRM and ERP	Posted records and an audit log	112 ms
02 Document parsing	PDFs, scans and phone photos	Validated JSON fields	94 ms
03 AI agents	A request on any channel	A finished multi-step job	1.2 s
04 Forecasting	Transaction and sensor history	Forecasts and risk scores	50 ms
05 Internal search	A question in plain language	An answer with citations	612 ms
06 Monitoring and upkeep	Telemetry from every pipeline	Alerts, retraining and fallback	Continuous

Module 01 · Workflow automation

Join up the systems that were never built to talk

Email, CRM, accounting, ERP and cloud storage joined into background pipelines that retry on failure and record every step. This is what replaces the Zapier and Make scripts that snap the moment somebody renames a field.

Specifications for the workflow automation module
Connectors	Exact Online, SAP, Salesforce, AFAS, HubSpot, PostgreSQL and Microsoft 365.
Throughput	About 112 ms per transaction, measured end to end.
When it breaks	Retries sit behind a circuit breaker, so nothing is dropped in silence.
Runs on	A dedicated EU cloud region, or your own hardware.

REST and webhooks Exact / SAP

See this one running in production

Pipeline event 200 OK

{
  "event_id": "EVT-9041-SYNC",
  "source": "INBOUND_MAILBOX_INVOICE",
  "pipeline": "FARVEX_EXACT_ONLINE_BRIDGE",
  "actions_completed": [
    "Extracted VAT ID: NL88231920B01",
    "Verified against Dutch Chamber (KVK)",
    "Posted Ledger 7100 in Exact Online",
    "Dispatched confirmation to Slack #finance"
  ],
  "latency_ms": 112,
  "status": "200_OK_VERIFIED"
}

Module 02 · Document parsing

Scanned paperwork back as data you can post

Invoices, customs papers, bills of lading and multi-page technical tables come back as JSON your systems can post straight away. Median parse time is 94 ms and extraction accuracy is 99.98%.

Specifications for the document parsing module
Takes	PDFs, scans, multi-page TIFF, phone photos and hand-annotated pages.
Accuracy	99.98%, checked field by field against algorithmic checksums.
Languages	Dutch, English, German, French, Spanish and Mandarin.
Tax	European BTW and VAT rates are verified before anything is posted.

Table extraction No training set needed

How the parsing pipeline is put together

Parser output 200 OK

{
  "document_type": "DUTCH_BTW_INVOICE",
  "confidence": 0.9998,
  "parsed_entities": {
    "supplier": "Rotterdam Freight Terminals B.V.",
    "kvk_number": "24189201",
    "subtotal_eur": 18450.00,
    "vat_21_pct": 3874.50,
    "total_eur": 22324.50,
    "iban": "NL91ABNA0412894120"
  },
  "validation": "ALL_CHECKSUMS_PASSED"
}

Module 03 · AI agents

Software that sees a multi-step job through

Agents triage support requests, check vendor contracts, raise inventory orders and escalate anomalies. You set the ceilings, and anything above them halts for a person to approve.

Specifications for the AI agents module
Sequencing	Each step is planned, executed and written to an audit log, in order.
Guardrails	Above the transaction value you set, a human has to sign off.
Channels	Email, WhatsApp Business, Intercom, Zendesk and webhooks.
Closed without help	82% of routine requests finish in under five seconds.

Multi-step tasks Hard guardrails

Read the access and approval model

Agent run log 200 OK

{
  "agent_id": "DISPATCH_AGENT_AMS_04",
  "task": "ORDER_RECONCILIATION_EXCEPTION",
  "steps_executed": [
    "Detected delivery mismatch on PO #8812",
    "Queried supplier SQL database via API",
    "Identified carrier delay in Rotterdam Port",
    "Issued automated client notice with revised ETA"
  ],
  "escalated": false,
  "execution_seconds": 1.2
}

Module 04 · Forecasting

Transaction history, read forward

Past transactions and live operational streams become demand forecasts, maintenance schedules and fraud scores — early enough that you can still act on them.

Specifications for the forecasting module
Models	Time-series transformers, gradient boosting and autoencoders.
Scoring speed	Under 50 ms on every ledger transaction.
Accuracy	94.6% over a 90-day horizon on customer data.
Alerts	PagerDuty, Slack and SMS, routed by severity.

Anomaly scoring Demand planning

See a forecasting deployment

Risk scoring output 200 OK

{
  "transaction_id": "TX-49102-EUR",
  "fraud_risk_score": 0.03,
  "confidence": "VERY_LOW_RISK",
  "demand_projection_q4": {
    "predicted_volume_tons": 42800,
    "confidence_interval_95": [41200, 44100],
    "recommended_safety_stock": 3800
  },
  "action": "AUTO_APPROVE_TRANSACTION"
}

Module 05 · Internal search

Put a plain question to your own archive

Manuals, contracts, tenders and years of email become searchable in plain Dutch or English. Every answer arrives with a citation to the paragraph behind it, and no document leaves your infrastructure.

Specifications for the internal search module
Indexed	Contracts, manuals, tenders, SharePoint, network shares and mail archives.
Answer time	Under 900 ms for a cited answer, citation lookup included.
Citations	Every claim links to its source paragraph, so a reviewer can check it.
Documents stay	On your own infrastructure. Nothing is sent to a third-party model API.

Sub-second search Stays on your infrastructure

Read how documents stay isolated

Search response 200 OK

{
  "query": "Wat is de opzegtermijn in het Maasvlakte-contract?",
  "answer_language": "nl",
  "citations": [
    { "document": "CONTRACT_MAASVLAKTE_2024.pdf", "page": 14, "clause": "12.3" },
    { "document": "ADDENDUM_2025_Q1.pdf", "page": 2, "clause": "4.1" }
  ],
  "indexed_pages": 1840000,
  "response_ms": 612,
  "left_your_network": false
}

Module 06 · Monitoring and upkeep

Somebody is still watching after launch day

Every pipeline reports throughput, error rate, latency and model drift into your dashboards. When a model starts sliding or a connector drops, we hear about it before you do.

Specifications for the monitoring and upkeep module
Watched	Throughput, error rates, p99 latency and model drift, via OpenTelemetry.
Drift	Scored continuously. Past the agreed threshold we retrain and tell you.
On failure	Traffic falls back automatically, the circuit breaker holds the job, and it replays from the buffer.
Who answers	An on-call engineer is paged inside the SLA in your contract.

EU cloud or on-premise Automatic fallback

See the five layers and how each one fails

Observability metrics 200 OK

// OBSERVABILITY SNAPSHOT
Metric: latency_p99_ms = 114.2
Active-Agents: 32 worker nodes
Drift-Score: 0.002 (Nominal)
Fallback-Events-24h: 0
Audit-Hash: e3b0c44298fc1c149afbf4c8...

Which of the six fits the work you are doing by hand?

Forty-five minutes: we take one workflow, work out what it costs you today, and tell you straight whether automating it pays back.

Request a briefing See it running in production

Six systems, wired into the software you already run