interest 01
Graph theory & combinatorics
My home field: edge colorings, matchings, and the structure of graphs. I like problems whose answers can be checked mechanically.
~/svah-x · waterloo, on · math undergrad
I study mathematics at the University of Waterloo, in Combinatorics & Optimization. Outside coursework I do machine learning research, mostly on mathematical reasoning and what goes on inside models.
research / blk 01
interest 01
My home field: edge colorings, matchings, and the structure of graphs. I like problems whose answers can be checked mechanically.
interest 02
A transformer computes by passing one hidden state through a stack of layers, each reading from it and writing back. I am interested in how information moves through that stack, and in ways to measure this on trained models.
interest 03
Reasoning does not have to be written out as text. Some models work by applying the same small network repeatedly, refining a latent state until it settles. I follow this line of work closely and reproduce papers from it, usually on puzzles where difficulty is easy to control.
projects / blk 02
Research prototypes, training infrastructure, and one hackathon. Code is on GitHub.
An AdamW variant that adjusts its step size using topological summaries of the local loss surface, computed with GUDHI. A safety lock limits the adjustment when the geometry estimate looks unreliable. Benchmarked against plain AdamW on CIFAR-10.
A DreamerV3-style agent that learns to play Geometry Dash from pixels. Most of the work was infrastructure: a custom Gymnasium environment and a Windows-to-WSL bridge that keeps observations, actions, and logs aligned at 60 frames per second.
Fine-tuning pipelines for Dream-7B and GPT-OSS-20B on mathematical text, using 4-bit QLoRA, gradient checkpointing, and DeepSpeed. On the benchmarks I used, math accuracy improved by about 20% while VRAM use dropped by more than half, so everything fits on a single 16 GB GPU.
A hackathon project from LA Hacks. It asks seven questions, then generates a twenty-years-older version of you: an aged portrait, a narrated life story, and a chat with that person.
A real-time CFD lab in the browser: a lattice-Boltzmann (D2Q9) solver with momentum-exchange force measurement, validated against the canonical cylinder benchmark. Put a standard dummy, a car, a rocket — or anything you draw or upload — into the test section; set wind speed, air temperature and surface friction; read the measured drag coefficient off the balance and watch the Kármán vortex street shed.
courses / blk 03
Notes from when I was learning the material, organized into two courses. Free and code-first.
background / blk 04
I started with competition mathematics (Euclid, first place in British Columbia) and came to machine learning from the math side.
At Waterloo I study Combinatorics & Optimization. A lot of machine learning turns out to be graph and optimization problems underneath, so the degree and the research fit together well.
Before the current work I mostly did reinforcement learning and training-efficiency projects; the ones above are from that period.
education & awards
toolkit
contact / readout
Open to research work and internships. Email is the best way to reach me.
[email protected]click to copy