FEATURED PROJECTS

Software, AI agents, robotics, and hardware projects by Evan Gong.

012026

OpenKyrozen

Self-Learning AI Agent · Terminal-Native

A terminal-native, fully autonomous AI agent that learns from every interaction — it operates your filesystem, manages git, fixes bugs through a 6-step protocol, and runs 20 self-learning features continuously in the background to build its own knowledge graph.

26 built-in tools — file I/O, shell execution, web search, 15 git operations, semantic memory
20 background self-learning features — fact extraction, skill invention, strategy distillation
Multi-provider LLM support with auto-fallback — DeepSeek, OpenAI, Claude, Gemini, Ollama
6-step bug-fixing protocol: reproduce → diagnose → hypothesise → fix → verify → explain
Web UI + REST API, Docker, pip package, MCP integration, PWA

View on GitHub

022026

Kyrozen

Product Creator & Manager · Full-Stack Agent Platform

A production-grade "product creator and manager" agent platform — Kyrozen turns ideas into shipped products with an autonomous core, desktop app, browser extension, and web frontend. It coordinates planning, implementation, verification, and deployment in a unified, self-hostable system.

Autonomous product lifecycle — from idea to deployed product in one platform
Multi-surface architecture — desktop app, browser extension, web frontend, and Python core
Self-hostable backend with Docker Compose, Caddy reverse proxy, and Postgres migrations
Real execution planning with activity tracking and audit-grade verification
Production-ready packaging — releases, install scripts, tests, and CI/CD scaffolding

View on GitHub

032026

OverflowBar

macOS Menu-Bar Utility · Native Swift App

A fluid second row for the macOS menu bar. OverflowBar keeps a crowded menu bar tidy by moving selected status items behind one persistent arrow, then revealing them in a fast, native second row whenever you need them — with Liquid Glass on macOS 26+.

Native macOS design — Liquid Glass on macOS 26+, lightweight material on macOS 15
Live status-item capture via Accessibility APIs, WindowServer metadata, and ScreenCaptureKit
Respects safe areas, notches, multiple displays, full-screen spaces, and reduced motion
System controls stay safe — Wi-Fi, Battery, Siri, Control Center, and Clock remain visible
One-click reveal by arrow or hover; selecting an icon activates its original control

View on GitHub

042026

Good Samaritan

Autonomous Open-Source Contributor · Python CLI

A deliberately cautious AI experiment: an independent Python CLI that discovers small open issues on GitHub, prepares focused fixes, tests them, gets a separate model review, and only then may open a pull request — with explicit AI disclosure and a dry-run-by-default safety model.

Dry-run by default — discovery, cloning, analysis, editing, validation, and review without remote writes
Safety guardrails — rejects assigned/PR'd issues, security work, bot-prohibited repos, and large diffs
SQLite-backed journal, memory, and maintainer-preference learning for better future attempts
Disposable dependency environments and command sandboxing inside temporary workspaces
Automatic provider fallback (Gemini, Groq, OpenRouter, local OmniRoute) with cooldowns on rate limits

View on GitHub

052026

Kards AI

KARDS Game AI Simulator · AlphaZero Training

A core, headless rules simulator for KARDS AI training. It models the full World-War-II card game as deterministic Python code, then trains an AlphaZero-style policy/value network through self-play — no UI, no networking, pure reinforcement-learning research.

Complete KARDS rules engine — card loader, serializable game state, actions, turn flow, effect engine
Custom-handler framework for card abilities, keywords, countermeasures, and battlefield rules
AlphaZero-style training loop with policy/value network, replay buffer, and vectorized self-play
Windows training orchestration — D-drive workspace, CUDA isolation, dashboard, and graceful stop
1488-card unmodified kards.info catalog as the single source of truth

View on GitHub

062026

Hardware Agent Runtime

Embedded Hardware Runtime · TypeScript MCP

A local-first runtime that lets external AI coding agents safely compile, flash, observe, experiment with, and verify real embedded hardware. The first target is Arduino through Arduino CLI, with adapter contracts that preserve a path to ESP-IDF, Zephyr, RP2040, STM32, and MicroPython.

Safe hardware-in-the-loop workflow — discovery, compile, flash, serial diagnostics, verification reports
MCP server exposes the runtime to AI agents via STDIO with input/output validation
Immutable, attributable evidence — inference never overwrites observation
Deterministic experiments with persisted human-action pauses and safety analysis
Validated end-to-end on real ESP32 Dev Module hardware

View on GitHub

072026

CAD It Up

Text-to-CAD Agent · Python/build123d

A text-to-CAD agent with a self-correcting verification loop. It turns a natural-language part description into an accurate, parametric B-rep solid (STEP and STL) using build123d over the OpenCascade kernel, then measures the result against a structured spec and self-corrects via LLM feedback.

Pydantic PartSpec model — mechanical parts as data (holes, counterbores, bolt patterns, walls, gussets)
Deterministic build() mapping — LLM emits spec JSON, never arbitrary build123d code
0.01 mm measurement assertions — bounding box, bore positions, bolt patterns, wall thickness
Self-correcting loop — precise diffs feed back to the LLM to regenerate the spec on failure
Exports production-ready STEP + STL with material and tolerance reports

View on GitHub

082026

MoodStudy

Mood & Learning Analytics · Hackathon (Team b1t)

A hackathon-built web app that helps students log daily moods, track study hours and sleep, and receive AI-powered insights — pairing an immersive Three.js UI with a Supabase backend and a Hugging Face LLM for personalized recommendations.

Mood tracking with emotional tags, notes, and pattern recognition
Learning analytics — study hours, focus levels, sleep & habit monitoring
AI-powered insights and recommendations via Hugging Face LLM
Built-in Pomodoro focus timer, GPA simulator, and academic planning
Immersive UI with Three.js, GSAP scroll animations, and Lenis smooth scroll

View on GitHub Live Site

092024

Anti-Fire Drone System

Autonomous Fire-Response Drone · Hardware

An autonomous fire-extinguishing drone system for rapid response — onboard vision, real-time path planning, and coordinated fleet deployment across hazardous zones.

Onboard computer vision for fire detection and tracking
Real-time path planning across hazardous zones
Coordinated multi-drone fleet deployment
Rapid-response autonomous dispatch

View on GitHub

102026

Burney-PDF

Liquid Glass PDF Previewer · Windows Desktop

A Liquid Glass–styled PDF previewer for Windows, built in C#. Features a modern frosted-glass UI with Ctrl+G page-jump navigation and smooth document browsing.

Liquid Glass (frosted/acrylic) UI on Windows
Ctrl+G page-number jump navigation
Smooth, responsive PDF rendering and browsing

View on GitHub