Yolina Yordanova

AI, from research to deployment.

I'm an MSc student in Data Science & AI at TU Delft, with a BSc in Computer Science from the University of Twente. I've deployed AI tools in industry, worked on interpretability research, and started a company.

I'm most interested in machine learning, AI safety, and technology that works for the people using it.

Yolina Yordanova

Now

Projects2024 – 2026

Container load optimizer

Edelman, 2026. Deployed and used daily by the sales team.

Sales staff were planning container loads by hand in Excel, and the plans often didn't work in practice. I was asked to write a ChatGPT prompt for it. Instead I built a constraint-programming model, and a web app around it that turns an order spreadsheet into a validated loading plan.

How it works

Grouping pallets into rows reduces the 3-D packing problem to a 1-D one, which is what makes it solve in seconds. The loading rules were reconstructed with the people who load the containers, and a validator checks every plan against them. CPMpy on CP-SAT, FastAPI, hosted on Render.

3-D view of a packed shipping container with pallet rows at the back and loose boxes in the tail
A loading plan for one container.

Causal abstraction in BERT

MSc research project, TU Delft, 2026. Team of seven. Being assessed for publication.

We reproduced Geiger et al. (2021) and extended it to arithmetic, asking whether BERT computes x ± y internally before deciding its sign. I owned the model and training pipeline.

The main finding turned out to be about the method: our setup could not tell the two variables apart by construction. We wrote that up, with the fix, rather than report an ambiguous result.

Line chart of alignment per BERT layer; the curves for the result and its sign lie on top of each other
Alignment per layer. The result and its sign never separate.

Clear Ahead Labs

Founded Dec 2025. Ongoing.

Digital work has a cognitive cost that nobody measures. Clear Ahead Labs tries to estimate it from behavioural signals, like typing rhythm and task switching, while keeping each person's data private.

So far: a literature review, an open-source signal-capture layer, and interviews with potential users that reshaped the product.

Thursday, 8:04
Good morning.
High capacity today.
Estimated capacity78%
Pattern detected

Your focus is strongest from 9–11. Early interruptions tend to fragment it.

Protect 9–11 today. Leave low-priority work until after.

A morning briefing from the early prototype.

RISE Student Rocketry Team

University of Twente, 2024 – 2025.

A 40-plus-person student team designing and launching rockets. I led external affairs: partners, sponsors, and telling the outside world what the engineers were building.

Grid of photos from the RISE team: engineers at work, sponsor partners, launch preparations
The team, 2024 – 2025.

AI workshops at Edelman

2025 – 2026.

A workshop series and an internal newsletter on how AI works and how to use it well, for colleagues without a technical background.

Four slides from an introductory AI workshop
Slides from the first session.

Smaller projects

  • 2025 – 2026
    Vase volume prediction Edelman

    Estimating the volume of irregular objects from photos, with computer vision and regression.

  • 2025
    Shiftly

    A scheduling tool that turns staff availability into a balanced roster.

  • 2024
    Smart construction sites platform BSc graduation project

    A full-stack platform for construction-site sensor data, for ECOLOGIC Research.

  • 2025
    Web infrastructure electricity consumption BSc thesis

    How much electricity the infrastructure behind the web uses.

Work

Education

Writing

What scammers can teach us about using AI at work

What AI-powered phishing scams can teach the rest of us.

Substack, 3 Oct 2025

More at substack.com/@yolinayordanova.

Contact

Email is the best way to reach me: yolinab@gmail.com