Document AI, Data Engineering & BI

Databases, document extraction, pipelines and reports, built by a team that runs its own data products.

Most of our own products are data products. uahub.dk holds all 873,347 Danish companies from the CVR register, danadresse.dk serves the Danish address register through an API, and soge.dk searches millions of vehicle and company records. We built and operate each of them, so when we design a database or a pipeline for you, it comes with lessons we paid for on our own systems.

This page covers data work: getting data out of documents and websites, storing it in a database that answers questions quickly, moving it on a schedule and turning it into reports. Keeping business apps in sync with each other, such as a shop with a warehouse or a CRM with accounting, is integration work and lives under automation and integrations.

What we build with data

OCR and data extraction from documents

Invoices, delivery notes, signed forms and scanned letters still arrive as PDFs and phone photos. We turn them into rows in a database: text recognition first, then field extraction, then rules that catch what OCR gets wrong, such as totals that do not add up, a CVR number that does not exist in the register, or a date in the future. Fields the rules cannot confirm go to a person instead of being guessed. It pays off when someone in your office retypes the same kind of document every week. A small example from our own work: a client answered our 35 project questions on paper, and six scanned pages became a structured summary in the project repository.

Database design that survives growth

A database that works with a thousand rows can fall over at a million. uahub keeps every active Danish company, and what we learned there shapes our client designs: PostgreSQL accepts at most 32,767 parameters per query, so batch sizes are calculated from row width; some register IDs overflow 32-bit integers; historical values such as ownership shares need a validity period, or the sums come out at 200 %. On soge.dk, browsing 12.6 million vehicle records stayed fast only after we removed a costly sort and validated filters before they reached MariaDB. We pick PostgreSQL or MySQL/MariaDB, version every change with Alembic migrations and add a search index where people type.

ETL pipelines into a database for reporting

Here ETL means a scheduled job that collects data from a source, cleans it and loads it into a database built for reporting and analysis. danadresse.dk is our clearest case: it keeps about 3.96 million Danish addresses in sync with the official address data and rebuilds its search index every night. A rewrite took that rebuild from about 7.7 hours to around 12 minutes, and each night the job compares the count in the database with the count in the index. uahub loads a daily delta from the CVR register the same way. We build pipelines that can be rerun safely, log what they changed and stop on a surprise rather than load half a file.

BI dashboards and KPI views

A dashboard is only as good as the counting under it. We build KPI views on your own database: bookings, revenue, no-shows, traffic split into people and bots, search performance. Our booking platform bkg.dk has occupancy and source reports built this way. The full page explains the approach and the price.

Learn more →

Web scraping and data collection

When the data you need sits on public websites, we collect it on a schedule: an official API or feed first, plain HTTP next, a real browser only where a page demands it. We keep request rates low, check the legal side first and watch every source for silent breakage. Prices start at 2,000 DKK.

Learn more →

Predictive analytics, honestly scoped

Forecasts need history, and many small businesses have one or two seasons of clean data at best. So we start with baselines and change detection before any model. On our own servers we compare the last two hours of load with each machine's median for the past day; tested on real data from 12 servers, that rule would have flagged a traffic surge four hours before our fixed thresholds did. We compare like with like, the same hour on the same type of day, because comparing a Saturday with a Friday once gave us a false alarm. If your data can carry a forecast, we build one and show its error on past periods; if it cannot, we say so.

Document workflows: imports, exports and PDFs

Documents must also go out, not only come in: exports your accountant can open, PDFs for print, imports that accept what Danish banks and Excel actually produce. In JaRyk, our payment-reminder product, the CSV import had to understand amounts like 1.234,50 kr., dates like 1.10.2026 and Windows-1252 files saved from Excel, and the bank-statement import reads exports from Danske Bank, Nordea, Jyske Bank and Lunar. For a Danish construction company we turned a long work chat with photos into a searchable web archive and print-ready PDFs of 199 and 33 pages. We build the template, the validation and the routing, so files open correctly the first time.

Reporting automation

A report nobody has to remember to make. Every morning our own system writes a KPI file per website with Google Search Console data for more than 30 of our sites and payment figures from Stripe, and a digest reaches Telegram before the working day starts. bkg.dk can send a salon owner a daily evening report with bookings and revenue, formatted as 4.350,00 kr. in Danish and DKK 4,350.00 in English. JaKonto exports invoices, VAT and profit and loss as CSV files with the byte-order mark Danish Excel expects. We agree with you what goes in, when it is sent and to whom; nothing goes to your customers without your approval.

Data systems we have built and run

uahub.dk

A Danish business directory with 873,347 companies from the CVR register, stored in PostgreSQL and searchable through Meilisearch. Ownership claims are checked against the public owner register. A daily job loads changes from CVR; we learned to schedule and batch those index updates after they slowed searches in the early morning.

danadresse.dk

Our own DAWA-compatible address API, built when the public DAWA service closed. FastAPI, MySQL and Meilisearch behind 56 endpoints, bulk cleansing of files up to 50 MB as background jobs, and a free tier of 1,000 lookups a day. The nightly index rebuild now takes about 12 minutes.

soge.dk

Vehicle and company search for Denmark, with an API. The vehicle table holds 12.6 million records, and company pages show financial ratios such as solvency, return on assets and gross margin, computed from the companies' published accounts. Our own job portal jajob.dk looks up employers through the same company API.

KPI reporting across our own sites

One KPI file per domain every morning with Search Console and Stripe figures, read by an operations routine for each product and summed up in a single Telegram digest. Each change on a site is logged as a bet with a metric and a deadline; on lynbro.dk a number-lookup guide took weekly search clicks from 16 to 37.

Tools we use for data work

Python with FastAPI, SQLAlchemy and Alembic; PostgreSQL, MySQL and MariaDB; Meilisearch where people search by typing; Redis for caches and job queues; httpx for HTTP collection and Playwright only where a browser is unavoidable; reportlab for generated PDFs. Everything is open source, so you are tied neither to a licence nor to us.

Prices for data work

Database creation
from 3,000 DKK
Document processing (PDF/Excel)
from 1,000 DKK
Data collection and scraping
from 2,000 DKK
Connecting one data source
from 1,000 DKK
Analysis and report building
500 DKK/hour

The price depends on the number of sources, how messy the data is and how often it must refresh. After a free scoping talk you get a fixed quote; send your request through the contact form.

How a data project runs

1

Sample the real data

Before any design we load a sample of your actual files or exports and count rows, empty fields, duplicates and odd formats. Decimal commas, two date formats in one column and similar surprises show up here, not in production.

2

Design the tables and the checks

We write the schema together with the rules that guard it: unique keys, valid CVR numbers and postcodes, sums that must match. You see it as a short document with example rows before we build.

3

Pilot small, then scale on purpose

The first run touches about 20 records, compared by hand with the source, then a larger batch, then everything. A pilot does not reveal every limit: at uahub a batch of 400 hit one that 20 never reached.

4

Schedule, watch and hand over

Jobs log what they changed and alert when a run returns zero or loses a field. You receive the code, the database and a runbook, and we can keep it running on project support.

Frequently Asked Questions

Can you start from the data we already keep in Excel?

Yes, that is the usual starting point. We read the sheets as they are, including Danish decimal commas and dates, map the columns to proper tables and show you the rows we could not place. Once the data lives in a database, colleagues who prefer Excel can still connect to it and work as before.

PostgreSQL or MySQL, does it matter?

For most small businesses both work well. We choose PostgreSQL for heavier analysis and strict data types, and MySQL or MariaDB when your existing system or hosting already runs it. The schema, the indexes and the backups matter more than the brand. Both are free and open source.

How accurate is OCR on our documents?

It depends on scan quality and layout, so we measure it on a sample of your own documents before we promise anything. Every extracted field passes through rules, and fields that fail go to a person. We would rather ask you to confirm five documents a day than load wrong amounts into your books.

Who owns the database and the code?

You do. The database runs on your server or on one we operate for you, and the code is handed over with a runbook. Nothing is locked into a paid platform of ours, and you can move it to another developer at any time.

Can you forecast our sales?

Possibly. First we check how much clean history you have and whether a simple baseline already answers the question. If a model adds accuracy, we build it and show its error on months it has not seen. If the data is too thin, we tell you so and suggest what to start recording now.

⚡

Get an Instant Price Estimate

Use our interactive calculator to get an approximate price for your project in 2 minutes.

⚡ Calculate Price

Ready to Cross the Bridge?

Let's build your lightning bridge to digital transformation. From first consultation to live deployment — we move fast.

Get Started Today