Projects
Research tooling, applied machine learning, and a few things built in a weekend. Numbers below are measured, not estimated.
Voice in, settled payment out — targeting Myanmar's remittance corridor
A voice-driven cross-border payments prototype. Speech is transcribed with Groq Whisper, intent is parsed by the Claude API, and the transfer settles on Hedera with an on-chain record. Built end to end during the hackathon.
The demo video could not be played here.
Open the file directlyAll projects
Nine-constraint symbolic evaluator · F1-vs-N sweep
The tooling behind my UROP fellowship: prompt-condition sweeps across nine models, symbolic scoring of generated Prolog, and a grid-size stress test that catches enumeration masquerading as rule induction.
Users across 18+ countries
A Burmese literacy platform running on Cloudflare Workers AI. Multilingual NLP pipeline served from the edge, with content recommendation driven by Workers AI models.
micrograd, reimplemented from scratch
Rebuilding Karpathy's micrograd line by line — autograd, backpropagation, the lot — as preparation for working on transformer internals rather than around them.
Hazardous recall 1.00 · AUC-ROC 0.958
Near-Earth-object hazard classification over 6,136 NASA records. Five per cent positive class handled with stratified splits and class weighting; full ETL pipeline with MLflow tracking.
99%+ validation accuracy · 30 FPS · under 10 ms/frame
MobileNetV2 fine-tuned for human vs. non-human face detection on a webcam stream, with the OpenCV preprocessing path trimmed until it held frame rate.
Serverless · zero cold start
A REST API for automated code review on Cloudflare Workers, with Llama 3.1 prompt-tuned to return structured findings on bugs, style, and security instead of free-form prose.
Three solver architectures benchmarked
Constraint-satisfaction, probabilistic inference, and heuristic search compared on the same boards. Written from scratch — a small taste of the symbolic side of neurosymbolic work.
Dataset filters & CI/CD for a public research index
Filtering and continuous-integration work on the DICE Lab's public catalogue of autoformalization literature, used to triage papers for ongoing survey work.
Nothing matches that combination. Try clearing the search box.
More
Coursework, experiments, and half-finished ideas included.