sheet 03

AI that works on your data,
and a number that says how well.

I take on a small number of projects alongside my full-time work at Ctrl+Space Labs. Remote, in English or Greek.

01What I build

RAG on your documents

An assistant or internal search that answers from your own files and cites them. Chunking, embeddings, retrieval and evaluation done properly, on infrastructure you control.

→ Working system, deployed, with a test set that shows how often it's right.

Document extraction

Scanned PDFs, forms or archives into structured data. Vision models with a schema, validation rules, and a clear accuracy number per field.

→ Pipeline plus dataset, and an evaluation report you can read.

Fine-tuning and model choice

When a prompt isn't enough: a fine-tuned or local model for a narrow task, chosen by measurement rather than by the newest release.

→ Model, eval harness, and the comparison that justified it.

Agents with tools

Agents that act: call your APIs, fill forms, run scheduled jobs. Small and specific beats general.

→ Agent, its tools, and the logs to see what it did.

02How a project runs
  1. step 1

    Call

    30 minutes. What you have, what you need, whether I'm the right person.

  2. step 2

    Test set first

    We pick 20–50 real examples and agree what a right answer looks like. That's how we'll know it works.

  3. step 3

    Build

    Small steps, each one scored against the test set. You see the numbers as they move.

  4. step 4

    Hand over

    Running system, the code, the evaluation, and notes your team can follow.

03Not a fit

Chatbots that only need a system prompt, AI features that exist for a press release, or work that needs someone on-site full time.

04Contact

Tell me what you're working with and what should come out the other end.

giannis@g-yfantidis.grLinkedIn