ESF+ project, 2025 to 2028

AIKAA

Generative AI for the small, repetitive jobs that fill the week in Finnish micro and small companies, built around problems the companies name themselves.

Funding
European Social Fund Plus, co-funded by the EU
Role
Responsible for development
Status
Running. Three demos built, not yet public.

01 The problem

In a company of ten people, nobody is hired to look things up, retype orders or rebuild last month’s report. Someone does it anyway, in between the work they were hired for.

  • Looking a product up in four places
  • Copying orders out of WhatsApp
  • Rewriting the same report

The project’s aim is plain: take that kind of work off people, so there is less rush in the day and more time for the part of the job that needed them. Whether generative AI is the right tool is decided one company at a time.

02 The programme

Thirty-two companies, four sectors

pilot companies
32
sectors
4
universities
4
months
28

  • Food industry 8 companies

  • Facility services 8 companies

  • Accounting 8 companies

  • Technology industry 8 companies

Led by Haaga-Helia University of Applied Sciences, with LAB University of Applied Sciences, Seinäjoki University of Applied Sciences and the University of Oulu. Funded by the European Social Fund Plus, co-funded by the EU, from November 2025 to February 2028. What works in the pilots is published openly as models any small company can use.

03 How a pilot runs

“Where does your week go?”

  1. 01

    The team

    A form first

    Each company starts with a questionnaire about how its work runs and where the time goes. It is the cheapest way to learn what to ask in person.

  2. 02

    The team

    One or two meetings

    The team visits once or twice and narrows the answers down to the problems that actually cost the company time. My colleagues lead this part.

  3. 03

    My part

    A working demo

    I build a demo around one of those problems, shaped to the company’s own kind of data, so nobody has to imagine how it would work.

    Example

    Photograph a sticky note

    an order in the system

  4. 04

    Together

    Show it, then decide

    The demo goes back to the company. They say whether it helps, what is missing, and whether generative AI is worth pursuing for that job at all.

  5. 05

    The project

    Write it up openly

    What holds up becomes a model other small companies can reuse, published at the end of the project.

04 The demos

Three demos so far

These are demos, shown to the companies and not public yet. Company names, logos and demo videos will appear here as each company agrees to share them.

01 Jewellery maker Demo, not public yet

Scan a piece, find its record

Thousands of pieces, many of them nearly identical: the same collection, similar shapes, small differences in size or finish. Checking a piece meant searching internal databases, paper product books and outside archives by hand.

  • Photograph a piece and get the three closest products, each with a confidence score and a sentence on why.
  • Add a product by pasting its web page. Details and photos come in on their own, and an admin picks which photos the system learns from.
  • Paper catalogues digitised, with a person approving every record, then searchable in a chat that shows its sources.

Demo video Coming soon

02 Bakery Demo, not public yet

Every order in one place

Orders arrived by WhatsApp, text message, phone, email and over the counter, across more than one site. The weekly roster was planned entirely by hand.

  • One order list for every channel. Paste an email, or photograph a sticky note, and the details are pulled out for you.
  • Staff send their shift preferences, and a language model drafts next week’s roster from them and the owner’s instructions.
  • A demand forecast from past orders, explained in plain language, that feeds the week’s production plan.

Demo video Coming soon

03 Reporting Demo, not public yet

Last month’s report, this month’s notes

Weekly reports, monthly updates, progress summaries. The content changes every time, the format barely does, and it is rebuilt from scratch anyway.

  • Define a type of document and give it good and bad examples to learn from.
  • Type rough notes, or say them out loud, and get a finished document in the same structure and style, ready to download for Word.
  • Ask questions of past reports, with the source report shown next to the answer.

Demo video Coming soon

05 Inside one scan

A photo in, three answers out

  1. 5%

    Clean the photo

    The background is cut away and the piece set on plain white, like the official product shots.

  2. 10%

    Identify it

    A vision model notes the type of piece, the metal, the style, the motifs and the finish.

  3. 35%

    Describe it

    It writes a few sentences about the piece, and those become a list of 3,072 numbers.

  4. 50%

    Shortlist

    Three searches run over the catalogue and are blended into one score. Eight go forward.

  5. 65%

    Second opinion

    The photo is set beside the official photos of the top five, to tell the same piece from a similar one.

  6. 100%

    Answer

    The best three come back, each with a confidence score and a reason.

Three searches, one score

Shortlist score
Signal Weight
Meaning of the description 70%
Keywords 20%
Text similarity, as a safety net 10%
  1. Not it 29% at most
  2. Possibly 30 to 69%
  3. Same piece 70% and up

The second opinion decides which scores are allowed. A piece that only looks similar cannot come back looking certain.

Built with production in mind, and still a demo. Real use will turn up edge cases, and those get worked through with the company before anything goes live.

06 My part

I am responsible for development on the project. The questionnaires and company meetings are mostly my colleagues’ work; I take what they find and turn it into something a company can try.

What I do

Build the demos

Turning a problem the team has found into a working demo, shaped to the company’s own kind of data. Three so far: a jewellery maker, a bakery and a reporting tool.

Work from the findings

The questionnaires and company meetings are mostly run by my colleagues. I work from what they find, and join in where the build needs it.

Build for production, call it a demo

Roles, logging and fallbacks are there from the start, but nothing is called finished until the company has used it and the edge cases have turned up.

Show it and listen

Each demo goes back to the company. Their reaction decides what happens next, which is a more honest measure than anything I could put in a report.

  • Python
  • FastAPI
  • React
  • TypeScript
  • PostgreSQL
  • pgvector
  • Celery
  • Redis
  • MinIO
  • GPT-4o
  • OpenAI embeddings
  • BiRefNet
  • Docker
Next project Voice to Value