Skip to content

I'm Warre Snaet, an AI engineer from Vilvoorde, Belgium.

In November 2026 I start as a physical AI engineer at Captic.

I graduated in AI Engineering at Howest, with distinction.

Work

  • Robot vision at Captic

    physical AI engineer, from November 2026

    Captic is a startup in Ghent that makes AI vision kits for factory robots. The software finds the next bag or box on a pallet and sends the robot a pick pose, or checks bags for damage. It runs on the machine itself, not in the cloud.

  • Medical imaging at 2Ai IPCA

    AI and software engineering intern, Portugal

    I wrote a Python pipeline that reads the measurements burned into echocardiography DICOM studies with OCR, extracts ECG traces, and finds the video frame each measurement image came from. That last step is an image similarity search, and it took most of the run time. I sped up the frame preparation and rewrote OpenCV's ECC alignment in PyTorch so it runs in batches on the GPU. Together, that made the search 14.9 times faster on a benchmark of 39 images, from 1182 to 80 seconds. A review tool lets a person check every value. It was an international internship, working with real clinical imaging data.

  • Apolloon 24-hour run system

    software engineer, paid client work

    It runs a 24-hour relay on three laptops on the event's own network, with no cloud. Each Electron laptop holds the whole race in SQLite. The laptops find each other through UDP broadcasts and elect a leader by majority vote with Raft. A change counts once two of the three laptops store it, so any one laptop can die without losing a lap. I built it alone.

Projects

  • AWS Fargate and Vault workload identity

    Terraform on AWS

    Terraform for private ECS Fargate tasks behind a public ALB, with subnets in two availability zones. In Vault mode, a Vault Agent sidecar logs in to an external Vault with the task's IAM role and hands the app the one secret it may read. It also sets up CloudWatch logs, CPU autoscaling from 1 to 3 tasks, and a deploy circuit breaker that rolls back. CI tests three plan variants against mocked providers and never deploys.

  • Financial AI agent

    school project, team of two

    A chat app for a simulated investment portfolio, made of six containers. A streaming FastAPI backend sorts messages into questions and trade commands. Questions go through RAG on ChromaDB, and trades go through eight MCP tools and need a confirmation. It uses PostgreSQL, a React frontend, and Qwen2.5 running locally through Ollama. I did the backend architecture, orchestration, persistence, containers and tests, and most of the frontend integration.

  • Azure ML lifecycle

    solo MLOps coursework

    A CNN that sorts photos into 15 kinds of sports ball. An Azure ML pipeline prepares each class in a parallel job, splits the data, trains the model and registers it. GitHub Actions sets up Azure, runs the pipeline, serves the model with FastAPI in Docker, checks it with a test prediction, and deletes the Azure resources afterwards.

  • Semi-supervised learning in Rust

    bachelor research, Rust and Burn

    A small CNN in Rust that classifies 38 plant diseases, trained with only 20% of the images labeled. Pseudo-labeling on the rest raised test accuracy from 86.06% to 94.90%. The release binary is about 26 MB and runs offline, and the model takes about 80 ms per image on an iPhone 12.

  • Event chatbot for Kortrijk Xpo

    school team project, lead developer

    A RAG chatbot for three trade fairs, built by a team of four for an outside client. I wrote most of it. A Scrapy pipeline collects the event websites. A .NET API embeds the content with Azure OpenAI, stores it in Cosmos DB, ranks it by cosine similarity, and answers with GPT-4. A Next.js dashboard shows the analytics.

  • Dataset query system

    Python client-server app

    Users log in and query a dataset of Los Angeles arrests through a PySide6 desktop app. The server handles each client in its own thread over TCP and runs the pandas queries in a worker pool. A separate server window shows who is connected and what they query, and can message one client or all of them.

Open source

  • Athas code editor

    contributor, 9 merged pull requests

    My largest change makes the editor, language servers and terminal stream large files. I also hardened how it handles untrusted input, added C# support, image diffs and dynamic proxy ports, and fixed Linux UI issues, resource and memory leaks, and a race condition in the language server handling. I work on the React and TypeScript side and on the Tauri and Rust side.

Writing