Distributed systems and ML infrastructure, implemented from the primary literature. I like the parts where a claim can be checked, so most of these repos ship a report whose every number came out of a run rather than out of a sentence.
Portfolio · USACO Gold Division · NLP paper submitted to NHSJS 2025
| what it is | the number that matters | |
|---|---|---|
| fluid-sim | Incompressible Navier-Stokes on WebGPU compute shaders, MAC staggered grid, five pressure solvers | Numerical viscosity measured at ν_num = 5.79e-4, 2.9% of the physical ν |
| ml-infra-platform | A Raft engine, an HNSW index and a training framework composed into one platform, driven from outside as read-only dependencies | First cross-layer bug: both components correct, every replica crashed |
| distributed-training-framework | Data-parallel SGD, ring all-reduce and parameter server, six invariants checked every step | 1,960 checked steps, 22 workers killed mid-step, worst gradient error 8.9e-16 |
| raft-chaos-testing | Per-link drop and delay injection against a real Raft implementation | 3 safety properties asserted per tick across a 600-seed sweep |
| vectordb-hnsw | HNSW from the Malkov & Yashunin paper, in NumPy. No FAISS, no hnswlib | recall@10 = 1.000, 20x faster than brute force at n=8000 |
| mcp-memory-server | Local-first persistent memory for Claude, HNSW over on-device ONNX embeddings | Zero network calls after the model is cached once |
Also: raft-kv-store (Raft from the paper, tested by killing servers mid-write over real TCP), mcts-code-reasoner (test-time tree search with a budget-discounted UCB1 and an RLIMIT sandbox), crisis-nlp-demo (attributions in closed form, so they are exact rather than sampled), and flightpath-scheduler (course scheduling as an NP-hard CSP, CP-SAT plus a dependency-free fallback).
From scratch means from scratch. The consensus protocol, the HNSW graph, the FFT, the multigrid V-cycle and the all-reduce are written here, not imported. Where a library would have done it, the repo says so and explains what implementing it taught.
The measuring code gets tested too. Three of the four bugs written up on fluid-sim's report were in the validators rather than the solver. A check that has never been observed to fail proves nothing, so the validators get deliberately broken to confirm they catch what they claim to.
Negative and partial results get published. fluid-sim's enstrophy cascade comes out at -4.529 against a theoretical -3, and the report says so and diagnoses why. Four of the six findings on distributed-training-framework's report are defects in its own harness, two of which would have produced a confidently wrong result. They are on the report rather than quietly fixed.
Python · C++ · Java · JavaScript · R · NumPy · WebGPU · asyncio