MW/26 CV

Founding Researcher at Humalike

Barcelona, Spain · 41.38° N, 2.17° E

I study how machines behave. Then I make them useful.

I’m Mateusz — a machine learning researcher and engineer working across language models, evaluation, and computer vision. I care about the point where a neat result has to survive contact with a real system.

See selected work Start a conversation

01 / Research

Making AI less generic.

My current work asks a deceptively simple question: can a model notice how a group behaves, instead of falling back to the same universally polite assistant voice?

Paper / 2026 Humalike

LoSoNA

A Benchmark for Local Social Norm Adaptation in Group Conversations

LoSoNA tests whether an LLM can infer an unstated local norm from a multi-party chat and apply it in its next response. We evaluated eight frontier and open-weight models across four prompting conditions — turning “reading the room” into something measurable.

8
models
4
prompting conditions
1
hidden local norm
Read the paper Explore the dataset
losona / scenario_042 context loaded
A

Ania I pushed the fix. It was the cache key after all.

M

Maks knew it. ship before it becomes sentient

J

Jan too late, it already requested code review

AI

You Great work, Ania! Your persistence really paid off.

02 / Field notes

Different problems.
Same instinct.

2026 — now

Founding Researcher

Humalike

Building the behavioral layer for AI that has to feel right around people. Current focus: social evaluation and adaptation in language models.

Barcelona

Feb — May 2026

ML Researcher

Sentinel Eagle

Computer vision for autonomous drone navigation when GPS is unavailable or unreliable.

Warsaw

2025 — 2026

Software Engineering — AI Intern

NVIDIA

Developed an evaluation platform that gave the latest language-model benchmarks one consistent interface, then worked on clustering related model errors.

Warsaw

2024 — 2025

Three kinds of applied ML

TSMC · MIM Fertility · Quant

LLM inference under hardware constraints in Taiwan, vision transformers for follicle assessment, and repository-level unit-test generation for a Singapore hedge fund.

Hsinchu / Warsaw / remote

03 / Build log

I like short feedback loops.

Research is one mode. A ticking hackathon clock is another. Both reward the same thing: finding the real problem quickly and making the idea hold together.

B / 01

Cockpit copilot

voice agents · aviation

A real-time “third pair of ears” that listens for dangerous situations and gives pilots timely suggestions through a conversational agent.

1st Warsaw Top 5 globally · ElevenLabs Worldwide Hackathon ↗

B / 02

AI Video Ad Maker

multimodal · agents

An automated pipeline that turns a product page into a scripted, voiced, animated, and testable short-form video ad.

1st + bounty Agentic Commerce & Best Voice AI · AIEngine 2025 ↗

B / 03

Last Minute

voice · planning

An event planner that calls venues, books restaurant tables, and handles the awkward coordination so people can just show up.

Top 3 finalist Warsaw.AI Hackathon 2025 ↗

Selected repositories

Built below the model layer, too.

Blob StoreC++20 · gRPC · Kubernetes distributed deduplicated storage ↗ Shared State MachineRust · networking shared data structures over the network ↗ Hearts Network GameC++ · client/server disconnect-safe multiplayer state ↗ BHL Churn PredictionCatBoost · Optuna tuned tabular ML pipeline ↗

04 / Scoreboard

A few useful constraints.

One night. A small team. A problem nobody fully understands yet. I tend to do well there.

  1. 01

    ElevenLabs Worldwide HackathonWarsaw winner · honorable mention, global top 5

  2. 02

    Goldman Sachs Engineering Hackathon1st place · Quant Challenge

  3. 03

    AIEngine Hackathon1st place · ElevenLabs Voice AI bounty

  4. 04

    AlgoTrade3rd in 2025 · 6th in 2026, 300 international participants

  5. 05

    ICPCUniversity of Warsaw representative · Central Europe Regional Contest

05 / About

The connecting thread is curiosity with a deadline.

I’m a Polish researcher and engineer, currently based in Barcelona. I started with algorithms and competitive programming, then kept following machine learning into increasingly different rooms.

At TSMC, the question was how models behave under memory and throughput constraints. At NVIDIA, how to evaluate them consistently and make their errors legible. At MIM Fertility and Sentinel Eagle, how perception can support decisions in the physical world. At Humalike, it is how AI behaves around people.

I earned my BSc in Computer Science from the University of Warsaw with a 5/5 diploma grade, and I’m now pursuing an MSc in Machine Learning after placing 2nd among 180 admission candidates. Outside work, I still enjoy the sharp constraints of ICPC, Codeforces, and building something from zero over one long weekend.

Open line

Have a hard ML problem
or a strange prototype?

mateusz.winiarek1@gmail.com