ML Engineer focused on computer vision for heavy industry — recycling facilities, scrapyards, and aluminum smelters. Currently Platform Lead at Visia, building multi-sensor perception systems (RGB, X-ray, CT, FLIR thermal) for hazard detection and material classification.
Background spans Physics → Data Engineering (Flask, SQL, AWS) → Computer Vision (behold.ai, active learning with PyTorch Lightning + BAAL, cleanlab-style data cleaning) → ML platform engineering. Now deep in agentic tooling, training infrastructure, and the full ML lifecycle from annotation to deployment on Jetson Orin edge devices. Based in Brooklyn, NYC. Writing occasionally at medium.com/@george.pearse.
| Project | Description |
|---|---|
| squeeze | Dimensionality reduction research platform in Rust - modular implementations of various techniques with benchmarking |
| tinify | Deep learning image compression research and implementation - exploring neural network approaches to image compression |
| bayesian_filters | Python Kalman filtering and optimal estimation library with Kalman, particle, EKF, UKF, and more |
| QDrant-NLP | Lightweight UIs and tooling for the QDrant vector database |
| ImageComposer | Package to simplify creating synthetic image datasets - specify backgrounds and foregrounds, the package does the rest |
| optillm-rs | A Rust port of optillm for use with AI coding agents |
I'm currently working on:
- Streamlit app for business intelligence (save SQL for transformations, plotly code for plotting).
- Quickdraw app via Streamlit (to try out some computer vision ideas).
- Active Learning via bagged ensemble disagreement.
- Lightweight UIs for the QDrant Vector Database (and all things QDrant because they build excellent tools) https://github.com/GeorgePearse/QDrant-NLP
- Multi-modal ML tools.
- How to use Modal without an exploding GPU bill
- The Missing Terminology of Data Centric AI
- How to train a model with MMDetection
- Data Engine Design
- Quick NLP Labelling with Bulk
