Projects

FLAGSHIP

Archery Apprentice

Production-Grade Cross-Platform Mobile Application

Kotlin Multiplatform Swift Jetpack Compose GitHub Actions JaCoCo CodeCov iOS Android

A comprehensive cross-platform mobile application for tracking archery scores, built with Kotlin Multiplatform for shared business logic and native UI implementations in Swift (iOS) and Jetpack Compose (Android). The project showcases production-grade development practices with extensive CI/CD infrastructure.

▸ 95%+ Code Coverage: Comprehensive test suite with JaCoCo integration and CodeCov reporting

▸ Automated CI/CD Pipeline: GitHub Actions workflows for automated testing, building, and deployment

▸ Cross-Platform Architecture: Shared Kotlin Multiplatform code for business logic with native UI layers

▸ iOS Native UI: SwiftUI implementation with seamless KMP integration

▸ Android Native UI: Jetpack Compose for modern, declarative Android interface

▸ Network Syncing: Real-time data synchronization across devices

▸ Production Best Practices: Dependency injection, repository pattern, clean architecture

Technical Highlights: This project demonstrates advanced DevOps capabilities including automated testing workflows, code quality gates, and cross-platform deployment pipelines. The CI/CD infrastructure ensures every commit is validated, tested, and ready for production deployment.

2024 • Mobile Development View Documentation →

ROGUELIKE

Rah6 — Roll a Hard 6

Balatro-Leaning Co-op Craps with Real-Casino Bet Fidelity

TypeScript React Vite Express WebSockets PostgreSQL Vitest npm Workspaces

A co-op craps game where a deck-builder chip/joker layer sits on top of a fully casino-faithful table. The whole game is built on a pure, deterministic TypeScript reducer over JSON-serializable state — the same property that makes online co-op, save/resume, replay, and deterministic tests one problem solved once.

▸ Pure Deterministic Engine: Seeded-RNG reducer with no I/O, no clocks — JSON round-trips losslessly by contract

▸ The Purist Test: Strip every chip and the table is faithful real-money craps — enforced mechanically by the type system, not convention

▸ Full Casino Bet Catalog: Every bet at real-casino payouts (pass/odds, come, place, buy/lay, props, hops, fire, ATS, repeaters)

▸ Roguelike Layer: Runs, antes, rarity-weighted chip shop, bet leveling, shooter rotation — the Balatro skin on top of the math

▸ Authoritative Online Architecture: Server-run reducer, per-bettor token auth, live lobby join + spectating over HTTP + WS

▸ 1000+ Tests: Deep engine coverage of casino payouts, determinism, and the chip contract, plus a real-browser Playwright smoke

Technical Highlights: A strict one-way layered monorepo (pure engine → authoritative server → React client). The chip contract is fenced at all three seams — a compile-time discriminated union plus two runtime guards — so the deck-builder chaos can never break the casino math.

2026 • Game / Systems Engineering View Documentation →

ENTERPRISE

AULMA

API Usage Logging & Monitoring Application

Apache Kafka Google BigQuery Spring Cloud Stream Java Microservices Event Streaming

Developed a scalable enterprise-grade application for monitoring partner API usage and asynchronously logging all API calls for analytics and reporting. Built using a cutting-edge tech stack to handle high-throughput data processing and real-time monitoring.

▸ Event-Driven Architecture: Apache Kafka for reliable stream processing and message queuing

▸ Scalable Data Storage: Google BigQuery integration for analytics-ready data warehousing

▸ Microservices Design: Spring Cloud Stream for building modular, independently deployable services

▸ Real-Time Monitoring: Live dashboards for tracking API usage patterns and anomalies

▸ Asynchronous Processing: Non-blocking event handlers for high-performance operation

▸ Partner Integration: Sponsored Search API and Play framework integration

Learning Experience: This project involved mastering several technologies I had never used before, including Apache Kafka, Google BigQuery, and Spring Cloud Stream. The challenge of building a production system with these new tools strengthened my ability to rapidly learn and apply cutting-edge technologies.

2023 • Backend System Learn More →

NSF

NSF - HPC Thermal Modeling & Prediction

Machine Learning for Energy-Efficient Computing

XGBoost Machine Learning Bash JavaScript Data Analysis

An NSF-funded undergraduate research project at California State University, Northridge, focused on developing machine learning models to predict thermal behavior in high-performance computing systems. Built an automated benchmarking pipeline and developed a highly accurate XGBoost model for temperature prediction across HPE BladeSystem c7000 infrastructure.

▸ NSF-Funded Research: Undergraduate research grant from the National Science Foundation

▸ XGBoost ML Model: Achieved 0.118 RMSE and 0.999 R² accuracy in thermal prediction

▸ Automated Pipeline: Developed Bash tooling for systematic benchmark execution on HPE BladeSystem c7000

▸ Web Dashboard: Built JavaScript-based visualization interface for real-time thermal data analysis

▸ Research Publication: Contributed to academic paper on HPC thermal management strategies

▸ Data Collection: Designed automated scripts for continuous sensor data capture across compute nodes

Research Impact: This NSF-funded research project provided hands-on experience in applied machine learning and high-performance computing optimization. The highly accurate thermal prediction model (R² = 0.999) demonstrated the potential for ML-driven approaches to energy-efficient computing infrastructure management, with implications for reducing operational costs in data centers and HPC facilities.

2022-2023 • Research Learn More →

NASA

IRP - Integrated Report Processor

NASA JPL Testbed Automation Tool

Python Pandas Data Analysis Automation NASA

A standalone report processor developed during my internship at NASA's Jet Propulsion Laboratory to assist instrument engineers with processing testbed reports from the Europa Clipper mission testing platform. The tool dramatically improved engineer productivity and report review efficiency.

▸ 70% File Size Reduction: Intelligent parsing reduced 600,000+ line reports by over 70%

▸ NASA JPL Deployment: Production tool used by mission-critical engineering teams

▸ Workflow Optimization: Significantly accelerated engineer report review process

▸ Data Processing: Advanced Pandas operations for efficient large-file handling

▸ Specification-Based Filtering: Intelligent extraction of relevant data based on mission specs

▸ Performance Optimization: Mitigated system issues caused by massive file sizes

Impact: This project remains my proudest achievement, contributing to NASA's Europa Clipper mission while solving real-world problems for instrument engineers. The experience of working on mission-critical systems taught me the importance of reliability, precision, and user-centric design.

2020-2021 • NASA Internship View Details →

GAME JAM

Dodging Game

JavaFX Game Development

Java JavaFX Game Development Game Jam

A simple but engaging game developed using Java and the JavaFX library during a 6-hour game jam workshop in 2020. Players control a character that must avoid randomly generated monsters moving across the screen. This was my first experience with both Java and the JavaFX framework.

▸ Rapid Development: Completed in 6 hours during game jam workshop

▸ First Java Project: Introduction to Java programming and object-oriented design

▸ JavaFX Framework: Learned GUI development and event handling

▸ Game Mechanics: Collision detection, random generation, player controls

2020 • Game Jam Play Demo →

HEALTHCARE

CareBridge

A cross-specialty clinical safety net that catches the gaps specialists miss.

TypeScript Turborepo Fastify + tRPC Next.js 15 Drizzle + PostgreSQL BullMQ + Redis Zod Claude API

CareBridge is a healthcare platform built to replace legacy patient-record systems, with an AI oversight layer that proactively detects dangerous cross-specialty clinical patterns. Every clinical data mutation flows through an event queue where a deterministic rules engine and an LLM review evaluate the patient's full picture, not just the latest chart entry. The system is engineered to stay within the FDA non-device CDS boundary with HIPAA-grade PHI encryption and a tamper-evident audit trail.

▸ Event-driven oversight: every chart change is reviewed across specialties asynchronously via BullMQ

▸ Hybrid detection: deterministic rules fire instantly, the Claude API handles nuanced cases

▸ Compliance by construction: non-device CDS posture, 7-year append-only audit log, AES-256 PHI encryption

▸ Versioned clinical prompts: golden-eval harness and a formal sign-off process guard prompt changes

▸ FHIR R4 interoperability: URL-form identifier namespace built for Epic/Cerner ingestion

▸ TypeScript monorepo: Turborepo and pnpm across shared packages, services, and Next.js portals

2026 • Healthcare Platform View Documentation →

DEV TOOLS

Chroxy

Remote terminal for Claude Code, Gemini, and Codex from your phone or desktop

Node.js WebSocket React Vite React Native Tauri Cloudflare Tunnel TweetNaCl

Chroxy is a self-hosted daemon that bridges remote clients to local AI coding CLIs over a secure Cloudflare tunnel. It gives every session a synchronized chat UI and full terminal view, with pluggable providers for Claude Code, Gemini, and Codex behind one WebSocket protocol. Mobile, desktop tray, and web dashboard clients all speak the same end-to-end encrypted contract.

▸ Pluggable Providers: Claude SDK/CLI/TUI, Gemini, Codex, and Docker variants behind one contract

▸ End-to-End Encryption: X25519 and XSalsa20-Poly1305 over an untrusted tunnel with forward secrecy

▸ Two Views One Session: markdown chat UI and xterm.js terminal on the same connection

▸ Container Isolation: sandbox, Docker, and git-worktree session isolation with resource limits

▸ Resilient by Design: supervisor auto-restart, auto-reconnect state machine, and push notifications

▸ Multi-Client: React Native mobile, Tauri desktop tray, and bundled web dashboard

2026 • Developer Tools View Documentation →

MED TECH

MedLens

Capture medical care, end to end.

React Native Expo TypeScript SQLite ML Kit OCR Expo Router Victory Native Jest

A local-first medical tracking app for patients and family caregivers. Photograph hospital artifacts like IV bags, lab printouts, and vitals monitors, then on-device OCR turns them into a structured, editable timeline that trends over weeks of treatment. Everything stays on the device, and clinical interpretation is firewalled out by design.

▸ Local-First Privacy: all records in on-device SQLite with no server holding medical data

▸ Tiered OCR Pipeline: on-device extraction by default, opt-in cloud tiers for higher accuracy

▸ Context-Aware Trends: coloring keyed to each metric's medically good direction, not just up or down

▸ SaMD Firewall: capture-only patient app deliberately separated from clinician-side interpretation

▸ Lock-Screen PII Rules: notification bodies stripped of patient identifiers by default, pinned by tests

▸ Store-Ready: feature-complete across five design phases, approaching App Store and Google Play submission

2026 • Mobile Health App View Documentation →

FAMILY

Letters to Loved Ones

Staying close to the people you love, in life and after loss

React Native Expo TypeScript Supabase Postgres SQLCipher Zustand NativeWind

A family app organized around four tracks for daily reflection, shared one-to-one and group prompts, and a grief-sensitive memorial experience. It is offline-first, private by default, and built so its mechanics soften automatically after a loss rather than punishing absence. The most delicate system, a server-enforced death-transition flow with an immutable audit log, protects users against premature or malicious memorialization.

▸ Four-track model: Personal, Pairings, Family Spaces, and decaying Memorial Candles

▸ Grief-sensitive design: streaks soften on loss, candles decay, absence is never punished

▸ Death-transition state machine: dual-path memorialization with cooldown, objection, and audit log

▸ Offline-first architecture: SQLCipher-encrypted local cache with last-write-wins sync

▸ Privacy by construction: row-level security, client-side encryption, signed-URL media, no ad model

▸ Purpose-built economics: bereavement freeze and a one-time Perpetual Memorial endowment

2026 • Mobile App View Documentation →

DEVOPS

Repo Relay

GitHub activity, relayed to Discord as threaded notifications.

TypeScript Node.js discord.js better-sqlite3 GitHub Actions SQLite Discord API ES2022

A standalone GitHub-to-Discord integration bot packaged as a composite GitHub Action. It posts one rich embed per pull request with an attached thread, routing pushes, CI status, reviews, and merges into that thread, and also relays issues, releases, deployments, and security alerts to dedicated channels. SQLite state keeps embeds updating across one-shot workflow runs.

▸ One-line install: an init command scaffolds the workflow automatically

▸ Threaded PR model: a single embed plus thread holds a PR's entire lifecycle

▸ Piggyback review detection: surfaces Copilot and agent-review despite GITHUB_TOKEN event limits

▸ Persistent state: SQLite WAL store survives self-hosted and cached GitHub-hosted runners

▸ Broad event coverage: 12 GitHub events spanning PRs, CI, issues, releases, deployments, and security alerts

2026 • Developer Tooling View Documentation →

MCP

Repo Memory

Persistent codebase memory for AI coding agents, over MCP

TypeScript Node.js 20+ MCP SDK SQLite better-sqlite3 Vitest ESM

An MCP server that gives AI coding agents persistent, structured memory about a codebase. It caches file summaries and re-reads only files that actually changed via SHA-256 hashing, cutting the tokens agents waste re-reading code they have already seen. Exposes 13 tools spanning summaries, dependency graphs, task memory, and token telemetry.

▸ Token Savings: roughly 3.6x summary-to-file compression, sustained from 10 to 200 files

▸ Stale-Proof Cache: SHA-256 hashing auto-refreshes changed files, never serves stale data

▸ 13 MCP Tools: summaries, change detection, dependency graphs, task memory, telemetry

▸ Cross-Turn Memory: task and session persistence survives context-window resets

▸ Multi-Language: regex summarizers for TypeScript, Python, Go, and Rust

▸ Measured Savings: built-in telemetry proves tokens kept out of the context window

2026 • Developer Tools / MCP Server View Documentation →

NAVAL

Sovereign Storm

Turn-based naval combat where every captain plans in secret and the sea resolves it all at once.

TypeScript Phaser 3 Colyseus Node.js Vite pnpm WebSocket

A browser-first, multiplayer, turn-based naval blockade game built as a pnpm monorepo. Captains plan moves and cannon fire simultaneously, and a server-authoritative engine resolves each round into a single deterministic replay so no two players ever desync. Ships five distinct modes spanning team PvP, co-op expeditions, a boss fight, and free-for-all.

▸ Server-Authoritative Netcode: one simulation ships a full replay, guaranteeing zero client desync

▸ Five Match Modes: Blockade, Voyage, Hunt, Free-for-All, and the themed Odyssey expedition over one engine

▸ Simultaneous-Turn Combat: secret planning plus deterministic resolution with ship-push, ramming, and shot-size damage

▸ Data-Driven Modes: a strategy registry lets new modes and themed expeditions ship as data, not engine branches

▸ One-Command Hosting: builds, serves, and opens a Cloudflare tunnel for zero-config co-op sessions

2026 • Multiplayer Game View Documentation →

MESSAGING

explAIn

Two-person messaging where LLMs help each party say what they mean.

Expo React Native NativeWind Supabase Postgres Vercel AI SDK TypeScript Turborepo

A one-to-one messaging app where every conversation has three voices: you, a private LLM advisor that helps refine your drafts, and a shared persona-driven mediator visible to both parties. Its non-negotiable rule is transparency: every transformation an LLM makes is visible to the people it affects. Personas range from a passthrough Moderator to a veto-capable Judge and a Dungeon Master that authors its own messages.

▸ Transparency contract: every LLM transformation is visible to the parties it affects

▸ Five mediator personas: distinct behaviors, not just prompts, switchable mid-thread

▸ Private advisor vs shared mediator: a load-bearing visibility boundary enforced by RLS and edge gates

▸ Cross-platform: one Expo codebase for iOS, Android, and web

▸ Provider-agnostic LLM layer: pluggable model providers, framework-free for a future bring-your-own-key variant

▸ Active/observer roles: structured-conversation asymmetry made a first-class primitive

2026 • Messaging / AI-Assisted Communication View Documentation →