Evan Huang

A complete underwater cave survey system I designed and built from scratch — custom hardware, bare-metal firmware, and a post-dive 3D analysis tool. The handheld device tracks a diver's position as they swim through a cave and renders a live map on a touchscreen display. Logged data loads into a web visualizer that reconstructs the cave in 3D.

A 90% 3D-printed toolchanger for FDM printers that achieves sub-millimeter repeatability with no sensors, no solenoids, and no electronics on the dock. The coupling geometry constrains tool position through contact alone — nine neodymium magnets provide clamping force. Designed for the Hypercube EVO with full Klipper macro automation.

A platform that automates commercial real estate due diligence — searching listings, analyzing properties, and comparing pricing across multiple data sources through a natural-language chat interface. Built across four services in three languages, coordinated by a shared protocol layer, and deployed on GCP with Terraform.

E-commerce site for a cave diving gear company I started — every product is designed from scratch in CAD and 3D-printed in ASA plastic. Customers configure colors on an interactive 3D model of the product before checkout. Built on Cloudflare Workers with Stripe for payments and Terraform-managed infrastructure.

Monitors flight prices and automatically detects mistake fares — flights priced far below market rate due to airline errors, which typically disappear within hours. Uses multiple independent detection methods with per-airline baselines, backed by a scraping system that handles sites blocking conventional HTTP requests.

Generates an animated, shareable year-in-travel recap from Polarsteps trip data — think Spotify Wrapped, but for your trips. Produces 12+ slides covering distance, countries, cities, and GPS traces with interactive maps, plus a parallel export system that renders the same slides as PNG images for sharing.

A crypto trading system that combines three independent analysis models — momentum, order flow anomalies, and on-chain liquidity — into a single signal, with built-in risk management that shifts to more conservative behavior when markets become volatile. Validated by backtesting across 100+ assets over 10 years of historical data.

Implements an alternative physics model for scuba decompression — calculating how long a diver can safely stay at depth before needing a slow ascent to avoid injury. The dominant industry model (Buhlmann ZH-L16C) treats gas absorption in tissue as simple exponential decay; this project models it as spatial diffusion and compares both across ~7,800 dive profiles.

AI coding tools can write Python fluently but have no built-in way to verify type correctness after making changes. This MCP server wraps Python's Pyright type checker so agents can check their own edits, see errors, and iterate — closing the feedback loop without waiting for a full CI run. Published on Docker Hub for drop-in use.

Applies the 'alternative data' approach used by quantitative hedge funds — satellite photos of parking lots, credit card transaction feeds — to 61.5 million torrent records spanning 19 years. Uses public piracy metadata as a demand proxy for the entertainment industry, then correlates activity against sector equities like Disney, Netflix, and EA.

A home server built to run local AI models, self-hosted services, and redundant storage — all managed as code. Hosts fine-tuned language models and image generation locally, eliminating dependence on cloud APIs and keeping all data on-premises. Provisioned and maintained through Ansible and Proxmox's API rather than manual configuration.

Home servers on residential internet connections have IP addresses that change without warning, silently breaking domain names. This daemon keeps Cloudflare DNS updated automatically — and specifically handles the case where a VPN connection would cause it to write the wrong IP, which simpler DDNS tools get wrong.

Streams a live fish tank camera from a Raspberry Pi to any browser with no cloud streaming service in between. The camera feeds directly into a video encoder, which pushes segments to a server over HTTP — the same delivery pattern used by professional encoders — and hls.js handles playback and error recovery on the browser side.

Editing a resume in a word processor means fighting formatting on every change and duplicating content for different roles. This system stores resume content as structured data and generates perfectly formatted PDFs on demand, with a web UI for drag-and-drop customization, role-specific presets, and AI-powered job matching to highlight the most relevant experience.