Search over your own documents
Ask questions of manuals, contracts, tickets or a wiki and get answers with links to the exact passages they came from.
e.g. internal knowledge base, support deflection, policy lookup
Applied machine learning practice
Models that do one job well, inside the tools you already use.
We are a small engineering practice. We take a concrete problem — finding the right clause in ten thousand contracts, reading invoices, answering support tickets with your own documentation — and build the part of it that a model can do reliably. Then we measure it, put it into production, and hand it over.
01 — Work
What we build
Narrow, testable systems. Each one starts from your data and ends with something your team can run without us.
Extraction and classification
Turn incoming documents and messages into structured records: fields from invoices, intents from emails, categories for tickets.
e.g. accounts payable, inbox triage, form processing
Assistants with a narrow brief
Drafting replies, summarising calls, preparing reports — with a person approving the result wherever a mistake would cost money.
e.g. support drafts, meeting notes, weekly summaries
Evaluation of existing AI features
Already shipped something with a model in it? We build test sets from real cases and measure how often it is right, and where it fails.
e.g. regression tests, model comparison, cost per answer
Private and on-premise deployment
Open models on your own servers when data cannot leave them, sized to the hardware you have rather than the hardware we would like.
e.g. regulated data, offline sites, fixed budgets
Data work that comes first
Most projects are mostly data: cleaning, labelling a few hundred real examples, and deciding what “correct” means for your case.
e.g. labelling guidelines, golden sets, pipelines
02 — Process
How a project runs
Short, with a decision point after each stage. You can stop after any of them and keep what was built.
Week 1
Scope
We look at real examples of the problem and agree on how success will be measured.
Weeks 2–3
Pilot
A working prototype on your data, tested against a set of cases you trust.
Weeks 4–8
Production
Integration, monitoring, cost limits and a fallback for when the model is unsure.
After
Handover
Documentation, the test set and a walkthrough, or ongoing support if you prefer.
03 — Terms
What you can expect
And what we will tell you not to do.
We will
- +Show numbers before promisesAccuracy on your own examples, not on a benchmark.
- +Keep your data where you want itNothing goes to an external provider without your written agreement.
- +Leave you able to maintain itYour repositories, your accounts, plain documentation.
- +Fix a price per stageAgreed before each stage starts.
We won't
- −Automate a decision that needs a personCredit, hiring, medical or legal calls stay with people.
- −Use a model where a rule will doIf a regular expression solves it, you get the regular expression.
- −Build a chatbot because it is expectedOnly if conversation is actually how your users want to work.
- −Lock you inNo proprietary platform between you and your system.
Tell us about the problem, not the technology.
A few paragraphs are enough: what happens today, where it hurts, and an example or two. We reply with questions, or with an honest view on whether machine learning is the right tool.
info@ai-driven.spaceReplies within two working days.