Prague, Czechia

Patrik Backo

AI engineer and independent builder.

Currently at Evangelist, researching vision-language models at CTU, and developing projects of my own.

Interested in contract projects and early-stage collaborations.

Have a project in mind?

LinkedIn Email More about my work CV GitHub

Work experience

Evangelist

Contract

AI Engineer

March 2026–present

At Evangelist, I build the data-processing pipelines behind short-form video campaigns. They collect posts and their engagement data, including video, descriptions, views, likes and comments, then turn that into campaign analysis and reporting.

My work includes analysing how view, like and comment counts change over time to spot suspicious engagement, checking brand safety and whether content matches the customer's campaign assets, and verifying creators' demographics: gender, age and country.

AI Researcher

March 2025–present

I study how the data used to train vision-language models affects the way they connect images and text. My research looks at models such as CLIP and SigLIP, with a focus on dataset bias and image retrieval: finding the right images from a text query.

ATG

Part-time

Deep Learning Vision Engineer

November 2025–February 2026

I worked on computer vision for industrial inspection, preparing datasets and developing deep-learning models for non-destructive testing.

MSD

Part-time

Data Science Intern

November 2023–October 2025

I worked on statistical analysis, machine learning and computer vision in Python, including deploying models with AWS and Databricks.

Siemens

Part-time

Student Software Developer

August–December 2022

I developed software in C# and Python and helped with Linux and network administration.

Education

Charles University

Faculty of Mathematics and Physics

Master's studies in Language Technologies and Computational Linguistics

2024–present

In progress · Thesis and state exams remaining

Bachelor's degree in Computer Science

2021–2024 · Completed 2024

Microcredentials

Deep Learning (2025)

Reinforcement Learning (2026)

Projects

Generating drum samples with a variational autoencoder

Bachelor's thesis · 2024

My bachelor's thesis explored drum synthesis using nearly 16,000 one-shot samples, which I collected and organised into nine drum categories. I trained variational autoencoders to reconstruct those sounds and generate new ones. The experiments went beyond random generation: they combined pairs of sounds in the model's latent space and varied individual samples along directions found in the data. This gave a way to explore variations of a drum hit, with the results documented through audio examples and analysis.

Source code on GitHub

SpeechCoach: spoken-English practice

Prototype

SpeechCoach puts a short speaking exercise in one place. You can generate an English sentence or enter your own, listen to a spoken example, then record yourself reading it. The app compares the speech recogniser's transcript with the target sentence and highlights words to review. It can also use the recogniser's confidence to flag uncertain words. A language model supplies practice suggestions for the highlighted words, and you can listen back to your recording before another attempt.

Source code on GitHub

Agentic Whiteboard: visual notes for coding sessions

Prototype

Keep the current design and its decisions visible while working with an AI coding agent.

Read more about Agentic Whiteboard

Agentic Whiteboard gives a coding conversation a visual companion: a page showing the current architecture, decisions and open questions. The plugin asks the agent to update that page as work progresses, keeping useful diagrams and their layout familiar. Earlier revisions are saved, so a previous design can be revisited after the conversation has moved on. The whiteboard can include interactive explanations, and each version is a self-contained web page that can be viewed locally without external services.

Claude Code hooks ask the agent to update the whiteboard during the conversation. Python helpers archive earlier revisions before edits and validate the new page; optional React components can be bundled into one HTML file. Codex follows a separate instruction-based workflow.

Source code on GitHub

Cottage trip registration and payment tracking

Archived

A reunion-trip website for signing up, checking payments and working out each person's refund.

Read more about the cottage app

This app was built for a cottage reunion with former high-school classmates. Participants chose one or two nights and received a private page with the correct amount and payment instructions. The organiser recorded incoming transfers manually and could see who had paid and how much remained to cover the cottage. If payments exceeded the target, the app calculated each paid participant's refund according to their number of nights, including exact rounding to the cent. It also tracked which refunds had been returned.

The app uses server-rendered pages and SQLite. Each registration keeps its original price. Refunds are calculated in integer cents, weighted by the nights booked by paid participants, with rounding remainders allocated so the total matches the surplus exactly.

Source code on GitHub

A little more about me

Outside of that, I'm usually climbing, hiking, running, or cooking.