Berlin. Incoming mathematics student, interested in machine learning systems built from first principles.
Portfolio: tsuskov.github.io
A complete language-model pipeline, written by hand without ML frameworks:
| Stage | Repository | |
|---|---|---|
| Tokenization | Cadmus | Byte-level BPE tokenizer — the merge-learning algorithm implemented from scratch |
| Training | Hephaistos | GPT/Llama-style model with hand-written forward pass and backpropagation, every gradient numerically verified; exports to GGUF |
| Inference | Talos | Minimal inference engine with an opt-in Metal GPU backend (2.8–3.7× the CPU path) |
| Deployment | Talos Forge | The full stack compiled to WebAssembly — live demo |
- raytracing — real-time CPU raytracer in Rust: multithreaded, BVH-accelerated, 4K export
- strange-attractors — interactive 3D visualization of the Lorenz and Rössler attractors, RK4 integration
- AloeGarden — SwiftUI focus-reading app for iOS with Home and Lock Screen widgets
- orbitale — hydrogen-atom orbitals (s, p, d, f) rendered as electron-density plots in the terminal
Learning theorem proving in Lean 4 (Mathematics in Lean) in preparation for university mathematics.



