Research engineering portfolio

Proofs, benchmarks, and systems with evidence close to the claim.

I am a mathematics undergraduate at the University of Washington. I build proof-oriented software, reproducible benchmark harnesses, certificate-shaped runtime tools, and data systems where the public claim can be checked against source, runs, and artifacts.

About

I turn mathematical ideas into software that can be checked.

My strongest projects sit between proof, algorithm design, and pragmatic engineering. I care about exact input formats, failure modes, reproducible commands, and clear boundaries around what has and has not been verified.

Current work includes Lean 4 formalization, normal forms over Euclidean and PID-like settings, sparse matrix benchmarking, certificate-oriented runtime monitoring, and data workflows where assumptions need to stay visible.

Lean 4 / mathlib Rust Python C++ Formal verification HPC benchmarking Algorithm design Reproducibility

Selected Projects

Six builds with traceable evidence paths.

Each project card says what exists, what it proves or measures, and where the current claim boundary stops.

Model evaluation pipeline

FlashDistill

A local Python/uv data factory for pre-Hyak math distillation experiments, keeping generation, normalization, verification, SFT packing, cost logs, reports, and data manifests in one CLI-driven workflow.

It stops at reproducible local gates and a Hyak handoff runbook. It produces JSONL artifacts and manifest hashes; it does not claim submitted GPU training runs.

Python uv LLM evals Schemas
Local artifact / public release pending

Civic data analysis

Seattle Transit Restoration

A Python analysis repo for schedule-based Seattle transit restoration work using King County Metro GTFS, archived Transitland GTFS, NOAA weather data, and ACS demographics.

The optimization claim is static-data connectivity analysis: routing modules, candidate generation, and MIP optimization. GTFS-RT scripts exist, but they are not the basis of the restoration result claims here.

Python Data analysis Civic tech
Open GitHub

Image tooling

agentic-color

A FastAPI + React/Vite MVP for image color variants. The backend accepts an upload and returns a schema-backed ImageAnalysis object with three StyleVariant presets.

The useful part is editability: each variant has explicit tone and color parameters plus a preview card. The analyzer still uses MVP/mock variants, so it is not described as a production ML grader.

Python Computer vision Image editing
Open GitHub

IoT case study

All-in-One Air Quality Monitor

A compact indoor monitor built with ESP32-C3, SEN55, onboard OLED, ESPHome firmware, and Home Assistant integration for PM, temperature, humidity, VOC, and NOx trend data.

The case study records the parts list, wiring table, firmware YAML, display behavior, and the black-screen debugging path that found the board's actual I2C pins.

ESP32 ESPHome Sensors
Read the case study