Dongjoo Lee이동주

Software Engineer

I released my first Android app in eighth grade, and the thrill of building a service I had imagined was reason enough to major in computer science. Designing and building software for complex domains at a company, in the military, and at my own startups, I’m growing into an engineer who understands the business and chooses the right technology for the situation.

Building

SaaS for running Christian communities

10,000 registered users · 2,300 MAU · KRW 000K monthly revenue

React Native · NestJS/Redis · MySQL/Prisma

  • Plan, design, build, and run the React Native iOS and Android apps and the web app
  • Built the NestJS, GraphQL, and Prisma/MySQL backend myself, and automated app and web releases with GitHub Actions CI/CD and OTA updates
  • Offloaded heavy work such as push notifications, image processing, and external API calls to Redis queues, and track performance and errors with AWS (S3, CloudFront) and Sentry
  • Served 100 concurrent users by running multiple Docker containers behind an Nginx upstream
  • Improve features and the user experience based on real feedback from our KakaoTalk channel and the user community group chat

AurisenCo-founder · CTO

— Present

Clinical recording and prescription analysis service

Raised KRW 20M from the Pre-Startup Package · KRW 20M from SNU BIG SCALE-UP

React Native · Electron · NestJS/Redis · MySQL/Prisma

  • Own product planning, UI/UX design, app development, and server infrastructure; shipped the iOS and Android apps to the app stores
  • Designed the NestJS, Prisma, and MySQL API along with asynchronous Redis/BullMQ workers
  • Implemented speech-to-text and structured LLM summaries for visit recordings, plus prescription OCR matched to medications through vector-embedding search
  • Set up AWS (EC2, S3, CloudFront) infrastructure, GitHub Actions CI/CD, and Sentry monitoring
  • Built an Electron desktop app that drafts SOAP notes for physicians in real time, now in pilot testing with several hospitals

Experience

ZUZU, Korea’s No. 1 shareholder management service, used by more than 13,000 private companies

React.js(TypeScript) · Django(Python) · RabbitMQ · Kubernetes/GKE

  • PaymentsDesigned a standalone payment module that makes new payment methods easy to add (credit cards, government non-face-to-face service vouchers)
  • SubscriptionsIsolated plan-based access control in a top-level layer so it never leaked into business logic, which let us absorb three pricing overhauls in a single year
  • NotificationsSplit a single queue by feature and applied late acknowledgement, building event-processing infrastructure with a 100% delivery rate and no lost messages even when servers go down
  • DeploymentStreamlined CI/CD, cutting deploy and test time from 14 minutes to 2.5
  • LegalBuilt automatic generation of share transfer agreements under the Korean Commercial Act, and a corporate registration product for bonus share issues
  • QualitySet up Storybook and React end-to-end testing infrastructure, and started Engineering Meeting, a knowledge-sharing session for engineers

Center for Teaching and Learning Innovation, ROKAF Education and Training Command

React.js(TypeScript) · Electron

  • Developed military intranet software and training simulators for military equipment
  • Designed and built an interactive simulator for Tactical Air Navigation equipment so first-time trainees can learn operating procedures without the extremely expensive real device; it was selected as the Air Force’s best training software
  • Recreated the look and feel of operating the real equipment with SVG interaction, HTML Canvas, and SVG optimization
  • Optimized rendering and bundle size to run on low-spec military intranet PCs, and designed a user-centered UI for novices

Awards

Bronze Prize (3rd) · Sogang Capstone Design Competition, Spring 2026 | Sogang University SW-Centered University Program

A system where a robot arm plugs in the charging connector so drivers never have to leave the car. I handled ACT imitation learning, the companion app, and Bluetooth integration between the robot arm and the app.

  • Trained autonomous control on data collected through teleoperation, using action chunking to keep errors from compounding.
  • Trained plug-in and plug-out models on 250 episodes covering normal parking as well as position and angle offsets, with three cameras (top, wrist, belly) and joint angles as inputs.
  • Built the mobile app that tells the arm to connect the charger, and integrated it with the ESP32 Bluetooth module.

Honorable Mention · Popularity Award · 2025 Sogang Convergence Technology Competition | Sogang University SW-Centered University Program

Our team of four won an Honorable Mention and also took the Popularity Award after placing first in the student vote. I designed a voice Q&A interface with STT, TTS, and VAD, and built an LLM chatbot that answers from evidence through RAG over board game manuals embedded as ‘expected question: answer’ pairs.

Gold Prize (1st) · 2023 Hyundai Mobis Mobility SW Hackathon | Hyundai Mobis

Proposed an architecture that sends a scene graph to a traffic control center, with the vehicles, pedestrians, and traffic lights detected by an autonomous vehicle’s cameras as nodes and their positions and interactions as edges, and built high-throughput messaging infrastructure with RabbitMQ and Docker Swarm.

Education

HCI Research

  • Observed how foreign users navigate the kiosks at Seoul Station and ran one-on-one interviews to investigate usability issues.
  • In a first field study that set our initial hypotheses aside, identified pain points in task flow, language guidance, and error correction.
  • Revised our hypotheses based on those findings, then ran a second field study to derive improvement directions and design implications.
  • Proposed kiosk UX improvements that account for how foreign users search for information, understand the language, and make decisions.
  • Ran an A/B test comparing the existing UI with an improved prototype among 50 members of the KGM Naver Cafe, followed by interviews with two of them.
  • Found that the prototype did not meaningfully improve perceived convenience, and analyzed the gap between users’ mental models and our initial hypotheses.
  • Combined quantitative preference data with qualitative interviews to lay out the limits of our prototype validation and where to go next.

Art Projects

15+ · Fantasy · Drama

Able to rewind the last 24 hours, Dave has spent years fixing every mistake and awkward conversation, stuck in a “perfect but repeating yesterday.” When a sudden downpour on a sunny day and his daughter’s wholehearted laughter hand him a moment no amount of rewinding can recreate, he realizes that life’s beauty lies in its unpredictability, smashes the hourglass, and embraces an unedited, imperfect present.

  • Scene 1Wide-angle lens · father’s eye level · over-the-shoulder shot. Past the father’s blurred shoulder, his seven-year-old daughter in a yellow raincoat plays in the rain at a rule-of-thirds intersection, while a row of trees leads the eye back to the father. Pouring rain and streetlight glow spreading across the wet pavement set a dark, melancholy mood.

p5.js · Three.js

Turn on the ceiling lamps one at a time and rotate the view left and right. Only when all three lamps are lit at full brightness can you finally tell that the object in the center is a matcha latte.

p5.js · WebGL

As I get to know new people, I grow more like them, and as my circle widens, I drift away from the relationships I already had.

Teaching

AI Problem Solving

Improving RAG Retrieval Quality

Board game assistant

Answer rate 60% → 96%

  • Chunking and embedding the manual as-is gave a 60% answer rate, and adding more data didn’t help.
  • I traced the problem not to a lack of data but to the semantic gap between conversational questions and the formal writing of the manual.
  • I built a dataset of ‘expected question: answer’ pairs from real community Q&A and questions anticipated from the rulebook, then switched to matching user questions against the expected questions.

Matching Medications on Prescriptions

Aurisen

Chose BGE-M3 by comparing top-1 / top-3 recall

  • Many prescription photos are blurry or only partly captured, so instead of trusting OCR output, I vector-indexed a standard drug database and matched by similarity.
  • Compared embedding models with six OCR misspellings of the target drug name and 11 easily confused drugs as candidates.
  • Drugs that only sounded alike sometimes ranked first, so embeddings now just generate candidates and the final ranking also weighs string-matching similarity.
  • Below 60% similarity, nothing is picked automatically; the top three candidates are shown instead, so a misread drug name never reaches medication guidance.

Clinical Visit Transcription Pipeline

Aurisen
  • Differences in pronunciation and regional dialects kept general-purpose STT from reaching the quality clinical records need. I post-process transcripts with an LLM to map terms to a standard medical terminology database, leaving uncertain spans marked [?] for the physician to resolve.
  • Built a tool for labeling transcription and chart errors as original–correction pairs, and used 31 pieces of clinician feedback as a golden dataset.

AI Agent Development Harness

Development environment

Productive agents with controlled cost and permissions

  • Kept per-domain context current in docs to eliminate the code exploration repeated every session, and delegated the remaining exploration and research to subagents to keep the main context clean.
  • Gave agents only minimal read access through MCP and required a person to enter credentials, keeping them out of the agent’s context.
  • Bundled build, test, and simulator verification into a single Skill so agents check their own results before reporting back.