IEEE WCCI 2026 Peer Reviewer · ML Researcher · Systems Developer @ Amazon Audible
Systems Developer at Amazon Audible, peer reviewer for IEEE WCCI 2026, and independent ML researcher with work currently under journal review.
I work on resource-constrained machine learning, compiler design, and production systems engineering. From migrating Player Services across AWS regions to building Edge Python, a single-pass SSA compiler and bytecode VM in Rust and WebAssembly (>240 stars, externally sponsored), and proposing uncertainty-aware classifiers for TinyML deployment validated on NASA IMS and SemEval.
Most of what I build is small, fast, and deterministic. Open to collaborations on tiny compilers, selective classification, and embedded ML.
Revisiting Rosenblatt Perceptron: Robust High-Entropy Classification via Uncertainty Margins Author of paper, introducing an uncertainty-aware linear classifier with adaptive abstention margin for TinyML deployment. ~1 KB memory footprint, 9 ms latency, benchmarked against Bonsai, FastGRNN, ProtoNN, and LSTM.
- Paper source: uncertainty-simple-perceptron.tex
- Implementation: uncertainty-simple-perceptron
Edge Python Single-pass SSA bytecode compiler and threaded-code stack VM for a sandboxed Python subset: NaN-boxed values, inline caching, super-instruction fusion, pure-function memoization, mark-sweep GC. Coverage-guided fuzzing; sub-200 KB WebAssembly module runs in the browser.
- Website: edgepython.com
- Live demo: demo.edgepython.com
- Source: edge-python
Edge Python Official Packages:
- Host Packages: Official packages that embed host-side bridge code (e.g., JS) and expose it to Python through the capability protocol. Includes modules like requests for networking and DOM bindings for browser interaction.
- Standard Packages: Official .wasm standard-library packages, where each capability is a Rust crate compiled to wasm32 against the wasm-pdk ABI. Hosts load the resulting .wasm over the standard plugin contract, no custom embedder, no Rust on the consumer side.
