Built an EEG pipeline that reads raw brainwave data and tries to classify motor imagery in real time. Used MNE-Python for preprocessing — bandpass filters, ICA artifact rejection, CSP features. SVM and CNN inference. Validated against PhysioNet datasets. Ongoing — the hardest part is keeping noise out.

Set up Mistral 7B running fully offline using llama.cpp and GGUF quantization. Mostly wanted to understand how quantization works and what you lose at 4-bit. Inference at about 12 tokens/sec on CPU. Useful for air-gapped environments.

Experimented with variational quantum circuits on Qiskit. ZZFeatureMap encodes classical data into quantum states, VQC trained via COBYLA. Hit 92% classification accuracy on the test set vs 79% with a classical SVM — interesting result though the dataset is small enough that I'd want to validate more before drawing conclusions.

Built a multi-node transformer QA system across 3 partitions with federated inference and ensemble voting. More of an architecture learning exercise than a production system — wanted to understand distributed inference at a small scale.

OpenCV + MediaPipe webcam pipeline with a custom CNN classifier on top. Runs at 30fps, all CPU inference under 40ms. Built it to learn how to build a training pipeline for custom gesture data.

Fine-tuned GPT-2 on biographical text data. After a few epochs it started producing convincing responses. More of an experiment in understanding fine-tuning mechanics than a practical tool.

ThinkCentre M73 Tiny running 10 Docker services — Immich, Jellyfin, Navidrome, and others — over Tailscale. Built to learn infrastructure, not because it was necessary. Useful for hosting personal media and experimenting with containerized services.