Get your data AI‑ready in 14 days.
We find, extract, clean and structure the data your AI product depends on — documents, databases, APIs, web sources — into a store your RAG, agents and evals can trust. Then we prove it on your own test questions.
a steel & mining group · a pharmaceutical manufacturer · one of Europe’s largest print houses · VIALink — see the work
Thanks — we’ll confirm your slot
- Within one business day you get a calendar invite for the time you picked (or the nearest free one).
- Before the call, send a short description or a sample of your data if you can. Not required.
- On the call we tell you whether the pilot fits — and if not, what we’d do instead.
From scattered sources to one store your AI can trust.
Most AI features fail in production for the same reason: the demo ran on a clean extract, the product runs on the real mess. We fix the mess.
# invoice_0417.pdf (OCR) ACME Sp. z o.o. NIP 525-23-44-178 Faktura nr FV/04/17 Data: 17.04.24 Razem brutto: 12 300,00 zł # crm_export.csv "Acme sp z oo",5252344178, "ACME Sp. z o.o. ",525-234-41-78, # support_ticket_8812.eml "…the Acme order from April was wrong…"
{
"entity_id": "org_5252344178",
"name": "ACME Sp. z o.o.",
"tax_id": "PL5252344178",
"invoices": [{ "no": "FV/04/17",
"date": "2024-04-17", "gross": 12300.00,
"currency": "PLN" }],
"tickets": ["8812"],
"source": ["pdf", "crm", "email"],
"embedding": "text-embedding · 1536d"
}
Sourcing
We map where the data actually lives and pull it in: file shares, your database, SaaS exports, public registries and the web — legally, with the source recorded for every record.
Extraction & cleaning
Parsing PDFs and scans, de-duplication, entity matching, normalising dates, amounts and IDs. The boring 80% that decides whether your AI answers correctly.
An AI-ready store
Structured tables plus a vector index (Postgres + pgvector by default, or your stack), chunked and tagged so retrieval can filter by customer, date or document type.
Evaluation
A test set of real questions with expected answers, run against the store. You see the score before and after — not a promise.
One use case, 14 days, a fixed price.
Small enough to start this month. Real enough to show whether the approach works on your data.
Scope and test set
We agree on one use case, up to 3 data sources, and 50 test questions with expected answers. These questions are the acceptance criteria.
Build the pipeline
Extraction, cleaning, entity matching and loading into the store. Daily short written updates.
Evaluate and tune
We run the test set, fix what fails and report the score per question.
Handover
- Pipeline code in your repository
- Populated store and the index
- Evaluation report and the test set
- A 60-minute walkthrough for your team
- Payment
- 50% at kickoff, 50% on acceptance.
- Acceptance
- The agreed test set reaches the score we fix on day 2.
- If we miss it
- We keep working at no extra cost for up to 10 days. Still missed — you don’t pay the second half.
- Security
- NDA before access. Data stays in the EU or in your own cloud.
- After
- Keep it and run it yourselves, or we maintain it monthly.
Data projects our team has delivered.
Data integration and analytics for manufacturing, pharma and the public sector — the same groundwork AI features now depend on.
Portfolio and project management dashboards: project costs, budgets, timelines, technology readiness (TRL) and economic effect, pulled from the project and document systems into one view for stakeholders.
SAP PPM · SAP PS
Microsoft SharePoint
People-analytics dashboards built on the HR system: performance, recruitment effectiveness, retention and turnover dynamics.
SAP BusinessObjects
HR system integration
Moved reporting from static Excel files to interactive dashboards: cost and volume by production line, revenue, plan vs. actual — available on web and mobile.
Five geo-tagged dashboards combining VIALink service data with Census data across parishes and households, for planning and community programmes.
Some clients are described by industry rather than by name. Details and references on request.
Who this is for.
Good fit
- Seed to Series B AI startups whose product answers from customer or domain data
- Teams whose demo works but production answers are wrong or out of date
- Engineers busy on the product who need the data layer done properly, fast
- Products that need European data: Polish and CEE registries, documents, languages
Not a fit
- Training foundation models from scratch
- Data labelling at scale
- Projects without a concrete use case yet — start with the free call instead
- Defence and political-campaign work
Before you book.
Why a fixed price? What if our data is bigger than expected?
The pilot is scoped to one use case and up to 3 sources so the price can be fixed. We check the scope on the free call, before you pay anything. If on day 2 it still turns out bigger, you choose: a smaller scope at the same price, or a separate quote for the rest.
Who owns the code and the data?
You do. Everything goes into your repository and your infrastructure. We keep no copy after handover unless you ask us to maintain it.
How do you handle sensitive data?
We sign an NDA and a data processing agreement before we get access. We prefer working inside your cloud account. Otherwise data is processed in the EU and deleted after handover.
Which stack do you use?
Python for pipelines and Postgres with pgvector by default. We adapt to what you already run — Pinecone, Weaviate, Qdrant, BigQuery, Snowflake, Supabase and others.
What happens after the pilot?
Either your team runs it — the walkthrough and the code are made for that — or we keep the pipeline fresh on a monthly retainer. No lock-in either way.
Where are you based?
Warsaw, Poland. We work in English with teams in the EU, the UK and the US, and overlap with US East Coast mornings.