I work on the theoretical foundations of computer science — algorithms, linear algebra, probability, and machine learning and I build things by deriving them, not by importing them. If a problem can be solved by calling .fit(), I'd rather first understand what's happening inside the call.
Currently building — MIT-OCW-Portfolio
A self-directed study through six MIT OpenCourseWare subjects discrete math, linear algebra, algorithms, probability, machine learning, and AI search where every core algorithm is implemented from first principles: NumPy only
Why build it this way:
- A library call proves you can use a tool. A from-scratch implementation proves you understand what the tool is doing. Writing backpropagation by hand means deriving every gradient via the chain rule not trusting autograd to get it right.
- Every script verifies itself against ground truth, not just "the output looks reasonable." The neural net's hand-derived gradients are checked against finite-difference approximation. The from scratch SVD is checked against
numpy.linalg. Dijkstra is checked against brute-force relaxation. A* is checked against Dijkstra for optimality while counting the nodes it saves. - The proof is in numbers, not adjectives — e.g. a linear classifier fails at 54.7% accuracy on a task a 2-hidden-layer network solves at 100%, demonstrating why non-linearity matters instead of asserting it.
Added a test verifying that to_parquet / read_parquet correctly round-trips uuid.UUID objects through PyArrow. While iterating on it across review rounds, the test surfaced a real regression: on Python 3.14 / nightly builds, PyArrow was failing to cast FIXED_LEN_BYTE_ARRAY data back into UUID objects, silently returning raw bytes instead. Traced it to PyArrow's pandas_compat casting layer, filed apache/arrow#50312 with a minimal repro, and the fix was merged upstream a few weeks later — closing the loop from "this test looks flaky" to "found and reported an actual upstream bug."
Languages
AI, Math & Data Science
Backend & Databases
Systems, DevOps & Architecture
Show, don't tell. If I can't derive it, I can't claim to understand it. If the code can't be checked against ground truth, it isn't proof of anything.
"Price is what you pay. Value is what you get."
"Code is ink; documentation is the footnote."
