De Anza College · Computer Engineering

I build systems that fly safely.

Planners and safety filters for drones in PX4 and Gazebo, and software for the people around them. Heading to ICUAS and AIGP in 2027.

Planners in PX4 SITL
6
Hackathons
6
Competitions, 2027
2
Public repos
13

I'm an engineering student drawn to machines that have to think for themselves: robots, drones, and the AI behind them. I learn by building, so most of what I know came from projects, research, and hackathons.

How I got here

  1. Age ?

    First contact

    How it started: first computer, first code, first robot

  2. 2023 · age ?

    Hsinchu Senior High

    Hsinchu, Taiwan.

  3. 2024 · age ?

    CS Clubs Alliance

    Lead coordinator for Hsinchu's high school computer science clubs.

  4. 2024 · age ?

    First lab: NYCU ICAL

    Research intern. Drones, ROS 2, PX4.

  5. 2026 · age ?

    Planners that fly

    APF, HOLO-DWA and MPC in PX4 SITL, plus D-Guide, a voice-commanded guide drone.

  6. 2026 · age ?

    Hackathons

    DAHacks 5.0, 2nd place with BikeScape. Cal Hacks 13.0 with Orkio.

  7. Now · age ?

    De Anza College

    Computer engineering, transferring to a UC.

Research · Math 1A Honors · in progress

Control Barrier Functions × PX4

  1. 01

    The planner asks for a velocity straight at the goal, through the wall.

  2. 02

    A barrier function h(x) marks the safe set: everywhere h ≥ 0.

  3. 03

    The safety filter bends the command just enough to stay inside, and the drone slides along the edge.

  4. 04

    But PX4 tracks velocity with lag τ. When α·τ > ¼ the guarantee breaks and the drone overshoots. The project derives that bound with calculus and tests it in PX4 SITL.

h(x) = 0 h(x) ≥ 0 · safe set u_nom α·τ > ¼
Fig. — nominal command (dashed), filtered (ink), with tracking lag (rust)

Autonomy lab · PX4 SITL + Gazebo

Six ways to not hit a pillar.

Each one is a full ROS 2 node flying a simulated multirotor through the same kind of obstacle field.

Take the drone apart →

APF · depth camera

Artificial Potential Field

A 3D field: the goal pulls, obstacles seen by an OAK-D Lite depth camera push, and the drone follows the sum in real time.

How it works → github.com/blar-tw/APF ↗

APF: Newton — push the object.

WARP: Einstein — curve the space it lives in.

WARP · Riemannian geometry

WARP

Obstacles enter a conformal metric that puts every surface at infinite geodesic distance. Collision would need an infinite path. Closed-form, no solver.

How it works → github.com/blar-tw/WARP ↗

Selected work

Orkio diagram: conflict resolution, task delegation, human-in-the-right-loop and Git around a shared context

Hackathon · Cal Hacks 13.0 · 2026

Orkio

A coordination layer for AI coding agents on different laptops. Hooks snapshot each agent’s unfinished work to Git, agents share plans over MCP, and an orchestrator merges the snapshots and runs tests to catch conflicts before they reach main. In a head-to-head build, it took 2 prompts instead of 9 and finished in 10 minutes instead of 16.

TypeScript · Factory Droid SDK · SpacetimeDB · Claude Code hooks · MCP

devpost ↗ github.com/chuddrw/CalHacks26 ↗
De Anza Compass landing page: a pixel-art harbor with the line “You’ve done the work. See where it takes you.”

Web app · Live · 2026

De Anza Compass

A free transfer and degree planner for De Anza students. Upload a Degree Works audit and see what’s left for UC transfer majors (from ASSIST) and which De Anza degrees and certificates you already qualify for, including ones you never applied for.

Next.js · React · Prisma · PostgreSQL · Supabase · Chrome extension

deanzacompass.com ↗

Hackathon · 12 hours · 2026

Constellation

A campus social app for students in their first weeks. Swap cards by QR code, watch your circle grow as an Obsidian-style graph, and keep notes on people you meet. AI suggests the notes; you approve every line.

React · Hono · Postgres · WebSocket

Bikescape Safe route map of Cupertino Bikescape Green route map of Cupertino under trees Safe ↔ Green

Hackathon · DAHacks 5.0 · 2026

Bikescape

Two bike maps for Cupertino. Every street gets a 1–10 safety score from Street View photos and city data; routing runs Dijkstra on “felt distance,” and real crash records are used only to validate the scores.

Python · Claude Vision · OSM · Dijkstra

github.com/blar-tw/Dahacks ↗

Also built

O-Guide travel planner screen

Hackathon · Taipei

O-Guide

A Taipei travel planner that checks lodging against city open data, builds itineraries with RAG, and pushes weather and crowd updates through an AI agent.

repo ↗
D-Guide route waypoints in a terminal log

ROS 2 · drone

D-Guide

A voice-commanded guide drone: it geocodes your destination, turns Google Maps walking directions into GPS waypoints, and flies ahead of you.

repo ↗

Robotics camp · 2026

Micromouse workshop

Line-following on the BitRacer Pro Max at the TMIRC summer camp, programmed two ways: MakeCode on micro:bit and Arduino on the on-board MCU.

repo ↗

Windows · Python

Screen Answer

Press a hotkey and Claude reads the problem under your mouse, then shows step-by-step work with rendered LaTeX, code blocks and matplotlib figures.

repo ↗

Robots should be provably safe — then useful.

About

I'm a computer engineering student at De Anza College, planning to transfer to a UC. I work between control theory and software: obstacle avoidance and safety filters for multirotors in PX4, ROS 2 and Gazebo, and apps built fast at hackathons.

Python · C++ · PyTorch · ROS 2 · PX4 · Gazebo · Docker

Looking for research and internship opportunities.

Let’s talk.