No Python, no PyTorch, no autograd. A full CNN with hand-written convolutions, pooling, and backpropagation using only Eigen and OpenCV.

From tokenisation and one-hot encoding to skip-gram embeddings, positional encoding, LSTM, and attention — a complete bottom-up reconstruction of the ideas behind modern NLP.

Using spatiotemporal deep learning to predict the next frame of a flood sequence, without ever seeing the correct answer first.

What I learned from building a Deep Convolutional GAN from scratch — five hard lessons about normalisation order, hyperparameter sensitivity, and why you should monitor the discriminator first.

A from-scratch NEAT implementation that tunes its own colour space parameters to identify objects through smoke, no labels required.

A conditional GAN prototype that uses location coordinates as a conditioning signal, designed to overlay depth-aware scene generation on top of thermal camera contours.

Implementing and comparing policy gradient, actor-critic, and off-policy methods in PyTorch — from VPG through to SAC with PER.

A full-stack agentic AI platform running on bare metal — hybrid retrieval, ReAct reasoning, MCP tool calls, and what it takes to own the entire stack.

Building an autonomous quadcopter from scratch — simulation, deep RL, and a custom physical build with no flight controller.

Why I stopped treating the GPU as a black box — naive matmul, shared memory tiling, and what the 138× gap to cuBLAS actually means.

RMSNorm, RoPE, and Grouped-Query Attention hand-written from scratch. Real architecture numbers read from the safetensors index before a single line of code.

The SwiGLU MLP block, the full DecoderLayer, and the complete Model class. Then immediately discovering that 4B parameters at fp32 is 16GB before you've done anything.

The vision half of the project. A hand-written SigLIP ViT verified against the real checkpoint, and the Projector that maps 768-dimensional patch embeddings into Qwen's language space.

The full stack connected and forward-pass verified for the first time. Then the more valuable part: five wrong turns on the way to actually understanding what SigLIP's training does.

The first time any weight in this project has been updated by gradient descent. Flickr8k dataset, hand-rolled preprocessing, MLflow tracking, and a loss curve that actually drops.

A real autoregressive generation loop, a FastAPI endpoint for before/after comparison, and watching noise turn into "the dog" for an image of dogs.

Stage 2 begins: maze navigation, coordinate tokens woven into reasoning text, and a discovery about what Qwen pre-reserved in its vocab before we even asked.

LoRA from scratch, a 1.24GB optimiser cost hiding behind a gradient hook, the tail distribution lesson, and stage 2's first real training run.

Correct syntax, correct endpoint, zero repetition: but no stop signal, and a suspiciously generic path. Loss going down and the model learning the task are different things.

Three different mazes, one identical output. Measuring cosine similarity at each pipeline stage to find where the image's distinguishing information actually gets lost.