Models I've trained and the interfaces I've built around them. Each ML project has a case study.
Illustration of the LangGraph control flow, not a recorded run.
Retrieval-augmented generation
Upload a PDF and ask it questions, including about its tables and images. Each answer points back to the page it came from.
Python
LangGraph
Gemini
FastAPI
Streamlit
Generative models
A GAN that turns a segmentation mask into a height map, which I then render as 3D terrain.
Python
PyTorch
NumPy
SciPy
Embedded computer vision
A YOLOv5n model with attention that spots crop pests in real time on a Raspberry Pi. I quantized and pruned it so it would run fast enough on the Pi.
PythonPyTorch
Ultralytics
Raspberry PiReinforcement learning
I trained PPO, TD3 and SAC agents to walk in BipedalWalker-v3 and compared how each one learned.
PythonPyTorchWebsites I designed and built myself.
Design challenge entry
My entry for Zineps' public redesign challenge: a 15-page site for a shipping platform, with an animated route globe, a carrier comparison and a pricing estimator.
React
TypeScript
Vite
Unofficial design concept
A design concept for takeUforward, a DSA learning platform, with guided walkthroughs, a learner workspace and ⌘K search.
ReactTypeScriptViteDraw up to three digits, then lower the weight precision and see when it starts getting them wrong.
An example 27. Draw over the pad to try your own.
What the model sees: each digit cut out, shrunk to 28 by 28 pixels and centred, the way MNIST digits are.
It reads
?
Every guess the model considered
Squeeze the weights
Squeeze the weights and watch it change its mind. At 8 bits the network is a quarter of the size and reads just as well, and at 4 bits it still holds. At 3 bits it starts guessing, and at 2 it forgets almost everything it learned.