Retyping documents
Invoices, CMR, packing lists and contracts are read by a person and typed into a system a second time.
Live automation lab — not slideware
See the automation run before you pay for it.
Working pipelines for documents, incoming requests, knowledge bases and CRM. Open-source components, your data stays where your policy allows, and nothing leaves the system without a human signature.
- A human signs off before anything is sent
- Every answer carries file, page and confidence
- Runs self-hosted; open-source core
- Received
- Extracted
- Verified
- Matched in base
- Drafted
- Awaiting approval
- Sent
consigneeZhejiang Ruifeng Ltd0.98
incotermsFOB Ningbo0.94
gross_weight— missingask
total_amountUSD 14 260.000.99
The pipeline stops here. Nothing is sent, changed in CRM or invoiced until a person approves it.
The starting point
Where the hours actually go
Sorting the inbox
Requests arrive by mail, form and messenger. Someone decides what it is, who owns it and what is missing.
Hunting for the rule
The price, the condition, the exception — it exists, in a file nobody can locate fast enough.
Assembling the same report
Quotes and client reports are rebuilt by hand every week from the same four sources.
The model
Not a chat. A process with a stop valve.
Every demo on this site runs the same seven-state track. Deterministic work — sums, rules, validation — is code and SQL. The model only extracts, classifies, summarises and drafts. The gate is not optional.
- 01
Received
Mail, form, Telegram, upload. Raw file stored untouched.
- 02
Extracted
OCR and field extraction into a typed JSON contract.
- 03
Verified
Schema validation, arithmetic checks, confidence per field.
- 04
Matched
Rules, prices and conditions retrieved with the source fragment.
- 05
Drafted
Task, reply, quote or PDF prepared — still internal.
- 06
Awaiting approval
A named person reviews raw data, draft and sources, then signs.
- 07
Sent
Only now the mail goes out, the CRM changes, the file is delivered.
Run it yourself
Six modules. One track. No sign-up to look.
Each module works standalone or chained into the full run. Sample files are preloaded — bring your own only when you want to.
Request dispatcher
Free-text mail or message in — service, contact, deadline, priority, missing fields and a reply draft out.
Try itDocument extraction
Invoice, CMR, packing list or contract in — typed fields, arithmetic checks and per-field confidence out.
Try itAnswers with sources
Ask your own documents. Every answer returns file, page and fragment — or an explicit "not confirmed in the base".
Try itQuote & report builder
A table or a request card in — a branded PDF and an email draft out. Generation is one deliberate click, never automatic.
Try itCRM copilot
Thread, documents and status in — a summary, the next step, a task and a reply waiting for approval out.
Try itTable analyst
CSV or spreadsheet in — calculations, charts, flagged anomalies and a plain-language explanation out.
Try itDemos run on demand and sleep when idle. A cold module needs a few seconds to wake — the button tells you when it is ready.
Control
What the model is allowed to touch
Nothing critical is autonomous
Sending mail, changing a deal, issuing a document — all of it waits behind a named approver. The model can prepare; it cannot commit.
Everything is checkable
Original file, raw extraction, source fragment, status and confidence stay visible. The clean version never overwrites the raw one.
Sensitive data is masked first
A local layer removes identifying data before any cloud call and restores it afterwards. For the strictest cases the cloud path is switched off entirely.
Your key, your model
Use our limited demo key, plug in your own, point at Azure OpenAI or Bedrock, or run an OpenAI-compatible endpoint inside your network.
| Context | Mode | Where the key lives |
|---|---|---|
| Public demo | Platform demo key | Separate provider project, spend-capped |
| Closed pilot | Ephemeral BYOK | RAM of an isolated session worker, 30–60 min TTL |
| Regular cloud work | Encrypted BYOK | Encrypted store, decrypted only for the call |
| Sensitive / enterprise | Local connector | Never leaves your infrastructure |
An honest limit: a key you type into our backend has passed through our backend. We do not store it in clear text, never expose it to front-end code, never write it to logs or analytics, and use it only for the calls you allowed. For production we recommend the local connector, where the key never travels at all.
Track record
Shipped, not sketched
Freight document intake
Invoices, CMR and packing lists for China–EU shipments parsed into a typed contract with validation before they reach the operator.
Read the caseAnswers with citations
Retrieval over internal documents where an answer without a verifiable source is treated as a failure, not a stylistic issue.
Source on GitHubMemory layer for agents
Mnemostroma — an offline memory sidecar that keeps context for AI agents locally, with no data leaving the host.
See the productCards marked “reference implementation” are our own builds, not client work. Client deployments are shown with the client’s permission; where the name is under NDA, the scope and the stack are still described in full and can be walked through live.
Engagement
From a demo to something your team runs on Monday
- 01
Discovery
One short form and one call: the process, the volumes, the systems, the privacy rules.
- 02
Audit
Where automation pays back, where it does not, and what the honest failure modes are.
- 03
Pilot
One pipeline on your real documents, measured against how the work is done today.
- 04
Integration
Wired into your mail, storage, CRM and reporting, with approval steps where they belong.
- 05
Handover
Runbook, training, monitoring and the option to keep everything self-hosted on your side.
- 06
Support
Maintenance, new document types, model changes, capacity tuning.
Most components are open-source. You are paying for analysis, design, integration, adaptation and support — not for tokens.
Under the hood
Open components, replaceable providers
- Docker
- Python · FastAPI
- PostgreSQL
- Qdrant
- OCR
- Traefik
- Model adapter
- API for your agents
Questions people actually ask
FAQ
Will my documents be used to train a model?
No. Demo uploads are deleted on a short schedule and are never used for training. In a deployment, storage and retention are yours to set, and the strict path keeps everything inside your network.
What happens when the model is wrong?
A wrong field is a visible field: low confidence is flagged, arithmetic is checked in code, and the approval step exists precisely because extraction is never perfect. An answer without a source is returned as “not confirmed”, not as a guess.
Do I have to move to the cloud?
No. The same pipelines run on your server or on a VPS you control. The cloud model is one interchangeable component behind an adapter, and it can be replaced with a local endpoint.
How long does a first result take?
A pilot on one document type or one request channel is typically the first milestone. The audit tells you the honest range for your volumes before any build starts.
Can our own agents call these tools?
Yes. Registered users get a scoped API key and OpenAPI schema for the same endpoints the site uses. An MCP server over the same tools is on the roadmap.
What does it cost?
Audit and design are fixed-scope. Build and integration are quoted per pipeline after discovery. Most software components are open-source, so licence cost is usually not the line item that matters.
Next step
Bring one process. Leave with a straight answer.
Thirty minutes, no deck. Describe the process you want automated and you get a feasibility read, the honest risks and a rough shape of the work.