01 Projects

Selected
projects.

Five projects with a closer look at what they do, what I worked on, and how the parts connect.

Featured projects ↓ · More work below

01Robotics + vision-language AI

Language-guided humanoid manipulation

A simulated Unitree G1 that turns varied language instructions into grounded, validated manipulation actions inside MuJoCo.

  • Integrated a locally deployed Qwen2.5-VL-7B perception pipeline with OpenCV refinement.
  • Built structured planning, inverse-kinematics motion, contact-aware grasping, recovery, and visual verification.
  • Created reproducible scenario evaluation, JSONL evidence, and 33 automated tests.
MuJoCoPyTorchQwen2.5-VLOpenCVDocker
INSTRUCTIONLANGUAGE INPUT PERCEPTIONQWEN2.5-VL + OPENCVGROUND OBJECT + SCENE PLANNER G1 CONTROLIK → REACH → GRASPTRANSPORT → RELEASE VERIFYTRACK + INSPECT
  1. InstructionA natural-language task enters the system.
  2. PerceptionQwen2.5-VL and OpenCV ground the object and scene.
  3. PlannerThe grounded task becomes a manipulation plan.
  4. G1 controlThe humanoid reaches, grasps, transports, and releases.
  5. VerificationThe result is inspected and recorded as evidence.
Closed-loop embodied workflowInstruction → perception → action → evidence
02Embedded system

Wearable fall detection device

Firmware for an STM32L4S5 wearable that classifies falls from accelerometer and gyroscope data, stores events, and exports them to a phone through NFC.

  • Implemented a five-state detector with jitter rejection and a prolonged-inactivity check.
  • Wrote a four-sample X/Y/Z moving-average filter in ARM Thumb assembly.
  • Encoded fall count, uptime, and timestamps as an NDEF URI for the web dashboard.
STM32CARM ThumbIMUNFC/NDEF
Wearable fall-detection firmware pipeline Accelerometer and gyroscope readings from the STM32L4S5 pass through a four-sample ARM Thumb moving-average filter and a five-state detector. Confirmed events store a count, uptime, and three timestamps before an NDEF URI is read by a phone over NFC. SENSOR INPUT STM32L4S5 CORTEX-M4 YXZ ACC X/Y/Z + GYRO BSP SENSOR DRIVERS MOVING AVERAGE X / Y / Z · FOUR SAMPLES t−3t−2t−1t ARM THUMB ASSEMBLY FIVE-STATE DETECTOR 01NORMAL 02FREE FALL 03IMPACT 04CONFIRMED 0510 s LONG-LIE TIMING THRESHOLDS + JITTER REJECTION CONFIRMED EVENT FALL COUNT+1 DEVICE UPTIMEms STORED TIMEST1 · T2 · T3 NDEF URI PHONE TAP
IMUACC X/Y/ZGYROSTM32L4S5 · Cortex-M4
4-sample moving average
t−3t−2t−1t
ARM Thumb assembly
Five-state detector
  1. Normal
  2. Free fall
  3. Impact
  4. Confirmed
  5. 10 s long-lie
Timing thresholds · jitter rejection
Event recordCount · uptime · T1/T2/T3
NDEF URIPhone tap
IMU → filter → five-state detector → event record → NFC
03Autonomous robotics

Search-and-rescue robot

A TurtleBot3 system that maps an unknown maze, explores frontiers, interrupts navigation for heat detections, and resumes after handling a target.

  • Integrated ROS 2 Humble, SLAM Toolbox, Nav2, LiDAR, an AMG8833 thermal array, and GPIO actuation.
  • Extended a frontier-exploration node to manage heat-source goals and avoid revisits.
  • Tuned the real robot’s navigation stack and built isolated sensor and actuator test utilities.
ROS 2Nav2SLAMRaspberry PiThermal sensing
Autonomous search-and-rescue control loop A TurtleBot3 maps a maze with LiDAR and SLAM Toolbox, follows frontier goals through Nav2, interrupts exploration when the AMG8833 detects heat, actuates a launcher through GPIO and PWM, marks the handled location, and resumes exploration. SLAM TOOLBOX · OCCUPANCY GRID FRONTIER EXPLORATIONMAP + ODOMETRYSELECT UNVISITED FRONTIERNAV2 GOALRASPBERRY PI 4 AMG8833 · 8×8 THERMALAMBIENT-RELATIVEPIXEL DISTRIBUTIONLEFTRIGHTFORWARDCONFIRMEDINTERRUPT NAVIGATION ON CONFIRMATION GPIO / PWM LAUNCHERACTUATE → MARK VISITED → RESUME
SLAM occupancy grid
Frontier node + Nav2Map and odometry → unvisited goal
AMG8833 thermal interruptLeft · right · forward · confirmed
GPIO/PWM launcherActuate → mark visited → resume
SLAM → frontier goal → thermal interrupt → actuation → resume
04FPGA + digital logic

K-Map simplifier and visualiser

A Basys 3 FPGA interface for entering three-variable truth tables and moving between K-Maps, Boolean expressions, and gate diagrams in real time.

  • Built switch input, SOP/POS selection, canonical-expression generation, and mode-aware control.
  • Created the 96×64 OLED pixel pipeline, reusable character renderer, themes, and cell highlighting.
  • Integrated independently developed submodules through the top-level design.
VerilogBasys 3OLEDSOP/POSVivado
8 SWITCHES TRUTH-TABLE INPUT SW0m0 · 0 SW1m1 · 0 SW2m2 · 1 SW3m3 · 1 SW4m4 · 0 SW5m5 · 0 SW6m6 · 0 SW7m7 · 1 SWi MAPS DIRECTLY TO mi TRUTH TABLE Σm(2,3,7) SIMPLIFIER SOP POS MINIMIZED OUTPUT 96×64 OLED · K-MAP VIEW A \ BC 00011110 A=0A=1 m00 m10 m31 m21 m40 m50 m71 m60 m2 + m3 → A′B m3 + m7 → BC M2 · M3 · M6 · M7 → B′ m0 + m2 + m4 + m6 → C′ F = A′B + BC Σm(2,3,7) → TWO GROUPS → TWO TERMS

Example 1: minterms 2, 3, and 7 simplify to A prime B plus BC.

96×64 OLED · K-MAP VIEW

A \ BC00011110 A=0 m00 m10 m31 m21 A=1 m40 m50 m71 m60
m2 + m3 → A′Bm3 + m7 → BC

F = A′B + BC

  1. 01SW2, SW3, SW7
  2. 02Σm(2,3,7)
  3. 03Group adjacent 1s
  4. 04Render result
Switches → Gray-code map → minimize → OLED
05Game + creative software

UniQuest

A two-person NUS-themed platformer built in Godot 4, with original systems and assets across desktop, web, and Android.

  • Designed reusable character state machines, combat, weapons, progression, and a global game manager.
  • Built JSON persistence with migration support, mobile controls, customization shaders, and automated GUT tests.
  • Created original game assets including pixel art, background music, and sound.
Godot 4GDScriptGUTAndroidWeb/PWA

Play the web build

UniQuest game architecture and runtime flow Keyboard or touchscreen input drives a reusable character state machine. Movement, combat, progression and scene systems update the global game manager, persist through a migration-aware JSON save, and ship to web, Android, Linux and Windows. GODOT 4 · 2D LEVEL PLAYER + ENEMY ENTITIES SHARE THE CHARACTER BASE REUSABLE CHARACTER FSM IDLE RUN JUMP FALL ATTACK DASH GLIDE DEATH CONTROL LAYERKEYBOARDTOUCHSCREEN GLOBAL GAME MANAGERSCENES · LEVELS · BOSSESCOINS · ABILITY UNLOCKSGPA · HEALTH · DEATHSRUNTIME STATE UPDATED JSON SAVE · FILEACCESS{ coins, abilities, weapons,colours, audio, GPA }RECURSIVE DEFAULT MERGE → MIGRATE EXPORT TARGETSWEB/PWAANDROIDDESKTOP
Godot 4 · 2D level
Character FSM
RunJumpFallAttack
Dash · glide · death
Global game managerScenes · progression · combat · GPA
JSON save + migrationCoins · unlocks · colours · audio
Export targets
Web/PWAAndroidDesktop
Input → character state → game systems → JSON save → export

06 More work

More projects.

A shorter list of other things I have built.

2026

Vision-language model robustness benchmark

Built a reproducible Qwen2.5-VL evaluation pipeline, visual prompt-injection attack, image-only defense, linked experiment artifacts, metrics, and automated tests.

PyTorchQwen2.5-VLOpenCV
2025

CuddleCare

Designed core command architecture, parsing, dynamic help, overdue-treatment queries, defensive persistence, logging, JUnit tests, and PlantUML documentation for a Java CLI.

JavaJUnit 5PlantUML
2024

Robot car builds

Programmed and wired obstacle-course and maze-navigating Arduino robots using motor control, ultrasonic, IR, LDR, and mechanical actuation.

ArduinoCSensors
2023

JavaCS

Built a server-rendered archive for 30 ISC Computer Science solutions with dynamic routing, JSON-backed content, search, filtering, and mobile navigation.

Node.jsExpressEJS
2020—22

GARUDA

Developed a multipurpose Discord bot with modular commands, MongoDB-backed configuration, persistent giveaways, moderation, and community tools.

Node.jsDiscord.jsMongoDB

Experience and education are in the résumé.

Open résumé ↗