Quality isn't a phase I run at the end — it's how I think about building software. I design test systems with one consistent shape across every kind of test that matters.

A test automation framework with four engines — API, UI, performance, and chaos — built against Vendure, an open-source GraphQL e-commerce platform. Every engine implements the same three-method interface — prepare → execute → collect artifacts — so the orchestration layer, fixtures, and reporting are shared across all test types. Adding a new kind of test means implementing those three methods; nothing else has to change.

• Four engines, one interface — API tests in pytest over a GraphQL client, UI tests in Playwright with Page Object Model, performance via k6 wrapped as a subprocess, and chaos via Toxiproxy injecting real network faults. The base interface enforces the order; the engines don't know about each other.
• Chaos with real faults, not mocked responses — Toxiproxy injects latency and connection-level failures into the live system, so the tests verify how the application actually behaves under network stress rather than how its error-handling code looks in isolation.
• Deliberate fixture scoping — session-scoped shared clients for read-heavy paths, function-scoped customers for checkout tests so state never leaks between cases. No hardcoded IDs; checkout discovers shipping and payment methods at runtime, so the suite works against any Vendure instance.
• End-to-end pipeline — a pytest plugin records every run to SQLite, a FastAPI + React dashboard reads from that database, and GitHub Actions runs all 44 tests on every push with reports uploaded as downloadable artifacts.

A capstone research project building a lightweight Transformer-based system for American Sign Language gesture recognition. The system processes ASL video, extracts spatial–temporal pose and motion features, and applies a custom compact architecture — TRM-Micro — to translate signed gestures into text.

• Owned the end-to-end data pipeline — from gathering and preparing the ASL video datasets, through MediaPipe-based feature extraction, to data ingestion feeding the model for training and evaluation.
• Contributed to model development — assisted in building TRM-Micro, which adapts recursive-reasoning principles from recent Transformer research to the spatio-temporal domain of sign language.
• Benchmarking — TRM-Micro v3 matched or exceeded the ST-GCN baseline on the ASL-Citizen dataset (Top-1: 61.46% vs 59.52%) at roughly one-fifth the parameter count, establishing a strong accuracy-to-efficiency trade-off.
• Ongoing — preparing the work for publication and extending toward continuous, sentence-level ASL recognition.

Repository kept private pending publication of the research paper.