TIMOTHY HABASHY · MIT
I make invisible
structure legible.
AI, physics, and quantitative systems—from sparse sensitivities to stellar motion.
A SHORT ORIENTATION
Two disciplines.
One instinct.
I’m an MIT student pursuing S.B. degrees in AI & Decision Making and Physics, with a 5.0/5.0 GPA. I like problems that force code, mathematics, and the physical world into the same room.
Find the signal. Measure the uncertainty. Trace what moved.
That instinct travels from laboratory electronics to sparse algebra, machine-learning ensembles, simulation, and astrophysics—from the scale of a derivative to the scale of a stellar neighborhood.
Python · PyTorch · NumPy · SciPy · pandas · GraphBLAS
C · Julia · MATLAB · SQL · Linux · SLURM · Git
Electronics · oscilloscopes · optics · data acquisition · COMSOL
01
FUNDED RESEARCH · MIT SSRC · RESEARCH IN PREPARATION
Every answer should remember how it changed.
A dual-number-style semiring lets a sparse computation carry three things at once: its value, its lineage, and its first-order sensitivity.
Sparse cells stay sparse. The derivative travels only where the algebra says it should.
y = 7.240 + — ε
- VALUE
- 7.240
- DERIVATIVE
- pending
- PROVENANCE
- A₁₁ → multiply → reduce
02
FUNDED RESEARCH · CENTER FOR ULTRACOLD ATOMS · FERMI 2
Noise is only invisible structure.
I built and calibrated the electronics that let an optical trap hear its own fluctuations: photodiodes, high-gain amplification, power supplies, alignment, and a careful inventory of every source of noise.
The useful pulse was always there. The instrument had to become quiet enough to admit it.
03
PROJECT · PYTORCH REPRODUCTION · IN PROGRESS
Disagreement can be a measurement.
I’m rebuilding TabM and TabM-mini from first principles, then testing whether implicit ensemble disagreement can expose distribution shift without giving up the efficiency that makes tabular models useful.
Regression · binary · multiclass · multi-seed runs · calibrated disagreement · member pruning · automated logs
04
PROJECT · GAIA DR3 STAR MAPPER
A catalog became a moving universe.
ADQL data became 3D position and velocity, then leapfrog integration turned the stellar neighborhood into trajectories at 5,000 years per frame.
Gaia source TH-0027
DRAG FIELD · CLICK STAR · ARROWS ROTATE
Most Significant Scientific Plot
MIT Annual Gaia Hackathon05
PROJECT · MONTE CARLO OPTION PRICING ENGINE
A price is a distribution wearing one number.
A memory-safe engine for European, digital, Asian, and barrier contracts— with batched GBM, honest standard errors, confidence intervals, and the paths left visible.
| BATCH | PATHS | ESTIMATE | STD. ERROR | 95% INTERVAL |
|---|---|---|---|---|
| 01 | 5,000 | 11.782 | 0.2413 | [11.309, 12.255] |
| 02 | 10,000 | 11.777 | 0.1706 | [11.443, 12.112] |
| 03 | 15,000 | 11.926 | 0.1393 | [11.653, 12.199] |
ACCUMULATING
THE NEXT VARIABLE