Hi there. I am a freelance Software Engineer and host of the inside view podcast, where we talk about AI progress. I have previously worked with the number one french bank (working on a computer vision feature), hash.ai (working directly with the CEO on scaling), azmed.co (detecting bone fractures in X-rays with computer vision), ledr.io (building a Python/C API), FloydHub (wrote deep reinforcement learning blogs) and the Future of Humanity Institute (interned in AI Safety there). You can find some of my writing on FloydHub, Medium, the Alignment Forum and Twitter. I also sometimes mix and produce music.
Implemented a new NLP feature for a chatbot. Currently developing the training pipeline, API and models for a Computer Vision module.
AZmed is a company detecting bone fractures in X-rays using state-of-the art computer vision architectures. I pushed different ensembling of fine-tuned models to production, maximizing both sensitivity and specificity.
My main research contribution is an organ classifier, where I built a Graphical user interface to re-annotate data, re-annotated myself thousands of images from medical feedback, then automatically about one million medical images with 80% accuracy. This was then used to improve our pre-trainings.
Wrote a wrapper in Python for a low-level network library in C and Ada. I then deployed this Python library to run predictions on time-series.
Open-sourced code for the paper "How useful is quantilization for mitigating specification-gaming?" by Ryan Carey, accepted at a workshop at ICLR 2019. I worked on mathematical models for AI deception, work featured in the Alignment Newsletter.
Tracked my time using colors every 15 minutes for 3 years. It uses a Conditionally Formatted Google sheets, which will automatically color cells when you type a number, where one number corresponds to one activity. The killer feature for me has mostly been to be able look at the shape of the data in colors to see if I could identify some patterns.
Reproduced two experiments from Prefrontal Cortex as a Meta-Reinforcement Learning System by simplifying the observation and action space, bringing the training time from 112 GPU-days to 1 CPU-day.
Wrote the code from scratch for all of the models of "Reinforcement Learning: an Introduction", matching the peformance of the 47 experiments.
The repo includes the solution for all of the exercises, anki flashcards summarizing the core concepts in the book and my code for all of the experiments.