PIPEFINgabriel@pipefin.comSoftware studio
Messy data in. Working software out.
Spreadsheets, CRM exports, email attachments, someone’s API — we pull it all into one clean dataset and build the software that runs on it, hosted on servers we manage ourselves.
Tell us what you’re building — gabriel@pipefin.com
or see what we’ve built ↓
Fig. 01 — messy data in, one dataset out
Architecture & systems design
The load-bearing decisions, drawn before anything is built.
Data sync & AI pipelines
Data pulled from every direction and made to agree with itself.
Ideas into production
Your idea, shipped as a real product on real infrastructure.
Featured — Nordic Smart Track
01 / 04Hundreds of companies. One comparable dataset.
Sales data arrives from hundreds of companies, each with its own column names, categories and habits. AI matching reconciles them — names, quantities, revenue, categories, whatever the columns hold — and anything scoring below the confidence line is flagged for a human before it counts. The result is a dataset the whole industry can be compared on.
Visit nordicsmarttrack.com ↗Fig. 02 — the model’s confidence score decides
- 0 – 1 every AI match is scored on how sure the model is
- ≥ 0.50 auto-mapped — 80–90% of labels clear this line
- < 0.50 a person decides — the 10–20% that remain
- < 0.40 retried with a stronger model first
- 1.00 exact name match — no AI involved
- 1 seen before? answered from the mapping table, no model call
- 2 exact match on a taxonomy name → 1.00, no AI
- 3 otherwise the AI proposes a path in the tree
- 4 under 0.40 it is described, then retried with a stronger model
- 5 nothing it still can’t place is guessed — it is left out and queued
Programmatic SEO that survives AI search
02 / 04Generating thousands of pages is the easy part, and it is also how most programmatic SEO fails: near-identical pages with nothing actually on them. Google buries those, and an AI assistant will never quote one. So a page only gets generated when there is real data to put on it, it carries structured data so a machine can read what it is, and it answers the question outright instead of working up to it — because more and more, the thing doing the searching is an agent that lifts an answer rather than a person browsing ten blue links.
Enough pages, each with something on it
Every place, city and country in the database gets a page, built from the records that already power the app — amenities, ratings, who reviewed it. None is written by hand and none is empty, because a page only exists once there is something real to put on it. Through the spring the page count kept growing while Google worked through the backlog, and the two compounded. It now earns traffic in 34+ countries, including queries in languages the pages aren’t written in.
- Impressions
- ×17.8
- Clicks
- ×6.8
- Avg. position
- 11.7 → 8.4
Ranking was never the problem
These pages were already being shown hundreds of thousands of times a month, and almost nobody clicked. In July 2026 we rewrote the titles and descriptions, restructured every page to lead with the answer, added structured data so results could render as more than a blue link, and repointed the set at the queries people actually type. Impressions barely moved. Clicks multiplied — the same pages, finally being chosen.
- Impressions
- ×2.1
- Clicks
- ×7.1
- Click-through
- 0.16% → 0.53%
- Avg. position
- 33.5 → 24.3
Google Search Console, 6 months, indexed so the shape shows without publishing a client’s traffic. The /B comparison is 3 weeks either side of the July change, so it is a like-for-like window rather than a longer one collecting more of everything.
Selected work
03 / 04
Babyccino
Livebabyccino.app ↗A map of cafés and restaurants that actually welcome toddlers. Built because we needed it.
Screenshot — babyccino.app
Customer map, Tres Media
In productionEvery customer, segment and ad area on one map, so the team can see coverage, overlap and conflicts at a glance.
Screenshot — internal toolboth directions · continuous · no spreadsheet emails
Sync API, customer platform
Under NDACustom endpoints keeping two systems in step, both directions, continuously.
New product, 2026
Under NDASomeone’s idea, built for real — stable and secure from the start.
04 / 04 — Foundation
It all runs on our own servers.
Machines we rent and set up ourselves, not a platform that charges by the request. Nothing goes to sleep; the first visitor of the morning gets the same speed as the last one at night.
- IsolationOne customer’s data can’t reach another’s — Postgres enforces it, not a code review.
- SurfaceNothing is exposed to the internet except the front door.
- BackupsRestored on a schedule, so we know they actually work.
- CostA fraction of the managed-cloud equivalent.
outbound-only tunnel · zero open ports
our machine — firewall: deny all inbound
Who
Pipefin is a software studio based in Sweden, working with companies in the Nordics and the US. The people who design the system are the ones who write the code and run the deploys — nothing gets thrown over a wall halfway through.
We take on a small number of projects and stay with them. No handover deck, no phase two that never arrives.