An assistant that answers from your documents, and shows the page.
For companies and software houses dealing with documents, customer support or repetitive work, and for SaaS products that want AI features.
The problem
The answer is usually already there, in a contract, a manual or an internal rule. The time goes into finding it, reading it and copying its data somewhere else. An AI assistant only helps if you know how often it gets the answer right.
What you get
An archive or assistant working on your own documents, with a written measure of how reliable it is.
- Answers with a source: each answer names the document and the page, so anyone can check it.
- Documents classified, data extracted: the app recognises each document type and pulls out the fields you need.
- A number before you commit: the share of correct answers on real questions.
How it starts
A small pilot project at a fixed price. We take a sample of your documents and the questions people actually ask you. At the end you have:
- the assistant working on the sample;
- the share of correct answers, question by question;
- where the data lives, and an estimate of model usage costs;
- a proposal for the full project.
How it is measured
We agree the test questions and the right answers together, before any code. The score comes from those, not from examples I picked. The same test runs again after every significant change, so you can see whether the assistant is getting better. Data stays on the cloud you choose, or on models running locally when documents cannot leave the company. Model usage is billed as its own line, separate from my work.
After the pilot
The full project follows the scope written in the proposal. After delivery there is a monthly retainer: keeping an eye on the score, model updates, small changes and support with written response times. The person who built it maintains it.
Work
- AI document webapp for a company. It classifies documents, extracts their data and answers questions about them: hundreds of documents a month, often dozens of pages long, and hours of manual work saved every day. I built the webapp solo and worked with other developers on the AI services that classify, extract and answer. Delivered in 2 months instead of the 4 quoted. Read the case →
- AI engine for a risk analysis platform. Six areas assessed in parallel by agents, on a layer that works across several model providers. I led the platform rewrite.
- Offline legal assistant (prototype). Model and dictation run on the computer: no data leaves the machine.
- More work →
FAQ
What does it cost?
It depends on how many documents you have and what the assistant needs to do, so I'd rather give you a real price than a range. Write me a couple of lines about your case and I'll reply with the cost of the pilot: fixed, agreed before we start, no extras. We estimate the full project together from the pilot's results.
Do our documents go to an AI vendor?
Only if you choose so. Your own cloud or local models are both options, and we settle it in the pilot.
What if the score is low?
You see it before deciding on the full project, along with what would raise it. If it is not worth it for your case, I will say so.
Can it live inside our product?
Yes. The assistant can become a feature of your software, connected to your data.