Full-stack engineer for real-time, distributed systems.
From the ROS graph to the operator screen — and through every failure mode between them.

kwanho.dev  ·  Editor's Introduction  ·  LinkedIn  ·  Instagram  ·  kwanho0096@gmail.com

After wrapping up the Pickit 3D contract (Aug 2025 — Mar 2026), I'm full-time on two side projects:

A YouTube Music taste analyzer. The one that knows you've been listening to the same 12 artists for three years.

• Chrome MV3 extension scrapes your YouTube Music liked tracks
• The web app enriches each track through Deezer and Last.fm, then runs a Gemini analysis pass to extract taste fingerprints
• Renders a taste profile, an interactive artist map, and a five-mode recommender (song / genre / unheard genre / indie / mix) with Tinder-style swipe rating
• Auth.js + Google OAuth · Neon Postgres + pgvector for embedding-based recommendations · Cloudflare Workers via OpenNext · Cron Worker for scheduled enrichment jobs
• Next.js · TypeScript · Neon pgvector · Cloudflare Workers · Chrome MV3 · Gemini

A local-first AI workspace where every answer cites the source it came from.

• Multi-provider — Anthropic Claude · OpenAI · Gemini · Moonshot Kimi · local Ollama, one interface across all of them. Per-model token usage priced and reported as inference cost per run
• Agent mode — decomposes a task into steps, runs tools (web search, file reading, document analysis), and re-plans as results arrive. Human approval gate before any action executes
• Evidence engine — maps every generated claim to its source document and flags unsupported claims. Hallucination at scale isn't a caveat, it's a system failure
• Document ingestion — PDF (+ OCR fallback) · DOCX · XLSX · Markdown · CSV. Context-selection filter so only relevant content reaches the model
• Deployed over Cloudflare Tunnel with local/remote access split and request-origin checks
• TypeScript · Python · Fastify · React · SQLite · Docker · Cloudflare Tunnel

42 Seoul → SEA:ME · Wolfsburg → Robert Bosch → Pickit 3D → today

Started programming at 14, while writing math and science essays at a school where clarity of reasoning mattered more than correct answers. That discipline — define every assumption, test logic against edge cases, pressure-test against reality — has stayed.

3D-vision robotic picking in production. The system can't be paused between runs — everything has to work, recover, and report correctly while a robot arm is mid-pick.

• Delivered production-support features across React UI + API + ROS / streaming layers to unblock on-site operations
• Built a reusable Python WebSocket client with reconnect/backoff and message validation; used as the integration baseline across internal tools
• Built an internal log workflow — timezone-aware download → upload/convert pipeline → multilingual UI — that cut manual log handling and sped up incident triage
• Shipped a configuration UI for system/network settings with input validation and safe defaults C · C++ · Python · TypeScript · React · WebSocket · ROS1 · Docker · Git

V2X is safety-critical automotive communication. An interface bug is not a ticket — it's an incident. First time my data contract decisions had formal consequences.

• Implemented a V2X vehicle demonstrator for safety-critical scenarios, defining message/data interfaces and validating end-to-end behaviour in test runs
• Integrated communication modules and designed data contracts to decouple components and accelerate iteration
• Built a React-based HMI for an eBike safety prototype C · Python · TypeScript · React · ROS2 · Docker · Git

Four-module curriculum co-initiated with Volkswagen: embedded systems, autonomous driving, connected vehicles, and a capstone with automotive industry partners.

• Embedded — low-level C/C++ on ARM (Raspberry Pi, Jetson), bare-metal drivers, sensor integration, real-time constraints
• Autonomous driving — ROS-based perception pipeline: lane detection, sensor fusion, path planning. The CARLA lane-keeping project below came from this module
• Connected vehicles — CAN bus communication, AUTOSAR architecture fundamentals, V2X protocol design — the foundation that made Bosch immediately legible
• First time engineering for reproducibility and certification: edge cases matter when the vehicle is moving; a fix that works 99% of the time is a 1% liability C · C++ · Python · ROS · CARLA

No lectures, no curriculum. A specification, three peer reviewers, and a deadline.

• C from scratch — libft, ft_printf, get_next_line — then progressively harder systems: process control, IPC, file descriptors, memory management with no safety net
• Minishell: POSIX-compatible shell — lexer, parser, built-ins, pipe chains, redirections, signal handling. Requires a precise model of what the kernel does on fork, exec, and dup; anything vague shows up as a bug
• Philosophers (dining concurrency): mutex and semaphore solutions under strict timing and resource constraints. Learned where assumptions about thread scheduling order break — and why race conditions in tests are worse than in production
• CPP modules 00–09: OOP through templates, STL, exception handling — each module peer-reviewed, not auto-graded
• Peer evaluation model: every submission reviewed by three peers. Trains the discipline of writing code that is readable and provably correct, not just functional C · C++ · Unix · algorithms · systems

An AR overlay system that recolors traffic signals and road markings in real time for colorblind drivers.

• Transparent display output, camera input, YOLOv5 object detection — prototyped end-to-end in ROS2 / Gazebo
• Plugin-based Python architecture: each new sensor modality (depth camera, eye-tracker) is a self-contained plugin, not a codebase change. Eye-tracking was the first live plugin, feeding gaze data into the detection pipeline
• Led and presented the cross-functional team to 1st place at BCX 2024, competing against automotive and tech teams from Bosch, Mercedes-Benz, Volkswagen, and others across Europe Python · C++ · ROS2 · Gazebo · YOLOv5 · eye-tracking

• Trained a CNN on CARLA simulator camera frames to detect lane boundaries and output a real-time steering correction signal
• Architecture: preprocessing (resize / normalize / augment) → convolutional feature extractor → regression head. Evaluated across varied road geometry, lighting, and traffic
• Built the full data pipeline from scratch: capture loop, annotation tooling, train/val/test split. Discovered early that model quality is bounded by data quality before it's bounded by architecture
• Failure mode analysis: regression head amplified prediction noise at high speed. Traced to low-confidence frames outside the training distribution Python · CNN · CARLA

• Founded and led a 50-member student community at École 42 Seoul — members spanning every stage of the program
• Delivered three live events for 200–300 attendees: performance sets, open stage, student band showcase. Managed venue, equipment, scheduling, and rehearsal logistics
• Built the planning process the next cohort inherited: timeline templates, equipment checklists, contingency buffers, risk-response playbook

Languages — C · C++ · Python · TypeScript / JavaScript

Frameworks — ROS1 / ROS2 · React · Next.js · Node · Fastify

Infrastructure — Docker · Git · Linux / Bash · Cloudflare Workers · Cloudflare Tunnel

Databases — PostgreSQL (Neon) · SQLite · pgvector

kwanho.dev  ·  Long-form profile  ·  LinkedIn  ·  Instagram  ·  kwanho0096@gmail.com