What we build
Five kinds of models. None of them off the shelf.
Each category below is a starting point. Your model is specified around your material, your wording and the systems it has to fit into — then built and tested by hand in our Berlin studio.
01 · Document intake
Document intake models
Incoming post, scans and email attachments arrive mixed together. An intake model reads each document, decides what it is — an application, a delivery note, a complaint, a contract draft — and sends it to the right folder, mailbox or colleague.
The categories are yours, not ours. If your office distinguishes seventeen kinds of incoming mail, the model learns those seventeen.
- Reads Scans, PDFs, email attachments, photographed paper
- Produces Document category, routing target, confidence value
- Typical desk Post room, shared inbox, registry
02 · Data extraction
Data extraction models
Forms, delivery notes and recurring statements all carry the same fields in roughly the same places. An extraction model reads them and writes the fields you define into clean tables — ready for Excel, your bookkeeping export or your internal database.
You define the fields and the format. The model is evaluated against real examples from your archive before delivery, with the measured accuracy written into your documentation.
- Reads Forms, delivery notes, recurring statements, standard letters
- Produces Structured rows: CSV, Excel or a direct system import
- Typical desk Order processing, administration, bookkeeping support
03 · Drafting & summaries
Drafting & summary models
Long reports, meeting minutes and project files pile up faster than anyone can read them. A drafting model condenses them into short, readable texts — and writes routine replies and standard passages in your house style, because it is trained on your own texts.
Nothing is sent anywhere. The model runs on your machines or ours and only ever sees the material you give it.
- Reads Reports, minutes, correspondence, project documentation
- Produces Summaries, draft replies, standard passages in house style
- Typical desk Team assistants, project offices, management support
04 · Internal Q&A
Internal Q&A models
"What is our rule for …?" — the answer is usually somewhere in a handbook, a guideline or an old project file. A Q&A model answers staff questions from your own documents and attaches the exact passage it used, so nobody has to trust it blindly.
It answers only from material you approved. Where your documents are silent, it says so instead of guessing.
- Reads Handbooks, guidelines, project files, archived correspondence
- Produces Answers with quoted source passages and document references
- Typical desk Onboarding, internal support, quality management
05 · Review & routing
Review & routing models
Tickets and requests arrive faster than anyone can sort them. A routing model reads each one, tags it by topic and urgency, and forwards it to the team that actually handles it — with a suggested priority your staff can override.
The tagging scheme comes from your workflow. If a request does not fit any category, it is flagged for a human instead of being forced into one.
- Reads Tickets, request forms, incoming messages
- Produces Topic tags, priority suggestion, routing decision
- Typical desk Internal service desks, facility requests, IT tickets
Model classes
Three sizes we work with
The right size depends on the task, not on fashion. Most office commissions land in the compact or standard class. The figures below are indicative ranges, not promises — your written specification names the exact class.
| Class | Typical size | Context | Runs on | Best for |
|---|---|---|---|---|
| Compact | ~2–4B parameters | ~8k tokens | An ordinary office PC or small server | One clear, narrow task — intake, routing, field extraction |
| Standard | ~7–9B parameters | 16–32k tokens | A workstation with a mid-range graphics card | Most office work — drafting, summaries, mixed documents |
| Extended | ~13B+ parameters | 32k+ tokens | A dedicated server | Broad duties across many document types — internal Q&A |
Deployment
It runs where you decide
Every model is delivered with documentation, a version number and a handover session for your team. Where it runs afterwards is your choice — and you can change your mind later.
Request a modelIn your office
The model runs on your own machines. Material never leaves your rooms — the strictest option, and the one most Berlin offices choose.
In our Berlin lab
We host the model on studio hardware at Hauptstraße 47. You reach it over an encrypted connection; we handle upkeep.
In your data centre
For organisations with their own infrastructure, we deliver for your data centre or private cloud and brief your IT on maintenance.