About me

A slightly longer version of who I am, and what I've built.

Everything should be as simple as possible, but not simpler.

— Albert Einstein

01 About

I'm Joy Timmermans, an ai & computer vision engineer based in Hasselt, Belgium. I like building vision systems that leave the notebook and survive contact with real cameras, real latency, and real users.

I started in applied computer science and robotics at PXL, then spent four years at Secury360 building the less-glamorous parts of production ML: data pipelines, temporal models, edge-device software, and all the glue that makes demos become actual products. It cut false positives by 40%, dropped cloud costs by 25%, and proved to me that the model is not the most important part.

These days I split my time between LeftFields, where I work on AI systems for clients, and Padelize, where I joined as fractional CTO after an audit turned into a full rebuild of the core system.

I care about useful AI, boring reliability, and good datasets. The best line of code is the one that never gets written — keep it contained, keep it simple, and ship things that do not fall apart when the lighting changes.

Currently building

Padelize

I first met Padelize through my work at LeftFields. What started as a small technical audit turned into a much bigger conversation: the implementation they had paid for over the previous year had no real foundation. It was hard to run, hard to trust, and not something you could sensibly build a product on.

I rebuilt the core system in about a month by going back to first principles instead of stacking tutorial code on top of tutorial code. Field detection, temporal modelling, player projection, cloud pieces, deployment, and the data loop all became simpler, more stable, and much easier to reason about.

After that, Padelize asked me to join. I did not want to leave LeftFields, so we made it work another way: I became fractional CTO with 10% ownership, responsible for the technical direction and the computer-vision system behind the product.

1 month

core rebuild
10%

ownership
90%

field-detection accuracy

Padelize court field-detection visualization

02 Work

LeftFields

February 2026 — Present

AI Engineer · Hasselt, Belgium

AI engineering across client engagements — from dataset infrastructure to secure edge deployment.

Rematics

Built source-to-data-lake workflows that cut dataset preparation from weeks to hours, and used a hardware security chip to create a fully unsupervised provisioning system — eliminating 30–40 minutes of manual setup per device and saving roughly 8,000 hours a year in manual work. Hardened deployment across Yocto-based operating systems and NVIDIA DeepStream pipelines.

Daymaker

Designed and shipped an analytics dashboard covering traffic, product usage, and behavioural flows — turning scattered product data into reporting views that drove day-to-day product and operational decisions.

RapidFit

Did a full company wide technology audit to help optimise their engineering and buisness processes, then proposed changes that cut selected turnaround times by up to 60%.

Padelize

July 2026 — Present

Fractional CTO · Hasselt, Belgium

Started with a technical audit through LeftFields, then rebuilt the unusable core system in about a month and joined as fractional CTO with 10% ownership. Own the architecture and technical direction across the mobile app, cloud infrastructure, AI services, and deployment pipeline. Lead the computer-vision work: temporal models, field detection, tracking, homography-based player projection, annotation standards, dataset review, and the feedback loop from model failures back into labelling and training.

Secury360

August 2022 — February 2026

Machine Learning Engineer · Zonhoven, Belgium

Built and maintained the computer-vision data platform on FiftyOne for multi-million-sample datasets. Applied data-centric AI to cut false positives by 30%, then developed a temporal preprocessing method that pushed the false-positive rate down a further 40% while improving accuracy. Shipped Azure-based AI filtering that lowered monthly cloud costs by 15%, and owned the Linux-based edge-device software in production.

PXL Research

February 2022 — July 2022

Intern Researcher — Time Series & Edge AI · Hasselt, Belgium

Industrial energy-optimisation research. Built multivariate time-series models that forecast 24-hour power consumption with 85% accuracy, and evaluated the edge-deployment constraints of running it in real time on sensor data.

PXL Smart-ICT

July 2020 — May 2021

React Native Developer · Hasselt, Belgium

Built a React Native application that streamlined communication between healthcare institutions and workers.

NYBE

May 2022 — July 2022

Junior Developer · Hasselt, Belgium

Contributed to scalable Laravel applications and production web-development workflows.

03 Education

Woolf University

2026 — Expected 2027

Master of Science in Artificial Intelligence · Remote

Coursework: Machine Learning, Deep Learning, Statistics, Data Analytics, Computer Vision, Generative AI, Agentic AI, AI Programming, LangChain, MCP, and AI Ethics.

PXL University of Applied Sciences and Arts

2019 — 2022

Bachelor of Applied Computer Science — AI & Robotics · Hasselt, Belgium

Honors: Graduated with Honors
Relevant coursework: Machine Learning, Robotics and Path Planning, Big Data, Large Language Models, and Time Series Analysis.

Selected numbers

30%

False-positive reduction

Data-centric AI @ Secury360
2.5×

Real-time inference speed

field-detection pipeline @ Padelize
15%

Cloud cost reduction

Azure-based AI filtering logic @ Secury360
8,000+

Hours saved per year

Unsupervised device provisioning @ Rematics
90%

Field-detection accuracy

Complete rewrite @ Padelize
60%

Process turnaround cut

quotation flow redesign @ RapidFit

04 Selected Projects

Self-Driving Bot

Robotics & Optimisation

Built a ROS-based autonomous robot using neural networks for lateral control, sensor processing, real-time inference, and path optimisation. Earned the first perfect score, 20/20, in the course history.

View on GitHub →

Chess Assistant

Computer Vision

Built a computer-vision system that recognises chessboard states and recommends moves through a self-developed chess engine.

View on GitHub →

05 Stack

Languages: Python , JavaScript , C , C++ , C# , Java , Kotlin
AI & CV: PyTorch , TensorFlow , Tinygrad , FiftyOne , Weights & Biases , Pandas , CVAT , Ultralytics , Roboflow
LLMs & Agentic AI: RAG , LangChain , LangGraph , MCP , LlamaIndex , Ollama , Hugging Face , Qdrant , Whisper
Edge & Inference: ONNX , TensorRT , DeepStream , ROS 2 , Yocto , Embedded Linux
Video & Streaming: OpenCV , FFmpeg , GStreamer
Engineering & Cloud: AWS , Azure , Docker , Git , Bash , GitHub Actions , FastAPI , Django , React Native , Laravel , SQLite , MongoDB
Other: Unity , Gazebo , LaTeX , Figma

06 Community

FiftyOne Community

Helped set up the Voxel51 community Discord server and built a custom bot that automated community interactions.

Link →

FiftyOne Contributions

Contributed PyTorch device flexibility and CVAT organisation support to FiftyOne, improving device compatibility and enterprise annotation workflows.

Link →

Speaking

Selected as one of four speakers at the Brussels AI Meetup.

Link →

07 Languages

Dutch Native
English Fluent
French Elementary

08 Contact

If you want to talk about computer vision, edge AI, weird datasets, or a project that might be fun to untangle, email is the easiest way to reach me.

joy.timmermans@leftfields.be GitHub — @timmermansjoy LinkedIn — /in/timmermansjoy