Hyeongju Lee 이형주
Data & backend engineer · 29 months in industryFinishing my master's in March 2027 · Military service completed

From collection to serving,
I build data systems that run in production.

In 29 months at Vision Inside (assistant manager), I built and deployed data collection, loading and serving for a KISTI-commissioned project, and the backend of a genomic report system. Now I research 3D reconstruction from lens array images in my master's at Kyushu Institute of Technology, and I rebuilt the image-data pipeline for joint research with Hitachi and deployed it on site in the Netherlands.

The path data takes, and the tools I have used at each stage

Main languages: Python · SQL
  1. 01Sources
    • Shopping sites
    • Public data APIs
    • VCF · FASTQ
    • Machine-vision cameras
    • Game client
  2. 02Collection
    • Selenium
    • BeautifulSoup
    • REST API
    • PySpin
    • Distributed crawling
  3. 03Storage
    • MySQL
    • AWS RDS
    • SQLAlchemy
    • Split, chunked loading
  4. 04Processing
    • SQL
    • pandas
    • scikit-learn
    • OpenCV
  5. 05Serving
    • FastAPI
    • Java · Spring
    • PyQt
  6. 06Operations
    • Docker
    • Harbor
    • AWS EC2
    • Cloudflare
    • Linux
01 Systems

Four systems I built and deployed

Drawn in the order the data flows. Click a node to highlight the path through it; below the diagram you see what I did and what others did.

Built by me Shared with others Others External sources and users
Aug 2022 – Jun 2024Vision InsideOutsourced contract

KISTI-commissioned project · an AI-based platform for creating value from public R&D (later SMART K2C)

Data engineer · collection, loading and serving were mineDesign: KISTI researchers

Project design KISTI researchers · Some model design commissioned university research · My part built and ran the data layer on Vision Inside's side (collection, loading, serving), and turned analysis models into FastAPI services and admin screens

  1. 01Sources
    • What I did
      Selenium crawlers collected it subcategory by subcategory, and the products were de-duplicated by product ID (ASIN).
      Later
      Moved to the Rainforest API.
      Stack
      • Selenium
      • Rainforest API
    • What I did
      Selenium crawlers collected it category by category.
      Stack
      • Selenium
      • BeautifulSoup
    • What I did
      Collected data from the Public Procurement Service API on the public data portal and crawled the national procurement site.
      Stack
      • REST API
      • Python
  2. 02Collection
    • What I did
      Built and ran a collection setup that split the crawlers across several laptops and loaded the results into one remote database.
      Why
      The crawlers were spread out to avoid being blocked by the sites they collected from.
      Stack
      • Python
      • Selenium
      • BeautifulSoup
      • MySQL
    • What I did
      Built the collector for the Public Procurement Service API on the public data portal. Moved Amazon collection to the Rainforest API.
      Stack
      • Python
      • REST API
  3. 03Storage
    • What I did
      Loaded the procurement, Danawa and Amazon data into a central MySQL database.
      Stack
      • MySQL
      • SQLAlchemy
      • pandas
  4. 04Processing
    • What I did
      Model 3: keyword matching between NTIS projects and companies (SQL). Model 4: price-band K-means clustering (number of clusters chosen by the Dunn index). Model 5: a daily batch that mines required-quality keywords from reviews (KeyBERT, Sentence-Transformers, Mecab). SQL statistics on procurement prices, attributes and deliveries.
      Stack
      • SQL
      • scikit-learn
      • KeyBERT
      • Sentence-Transformers
      • Mecab
    • Others
      Some of the platform's models were designed by university research teams (Sungkyunkwan University, Kookmin University). Vision Inside owned the data and serving layers those models use.
  5. 05Serving
    • What I did
      Served the analysis models as FastAPI services and built the admin screens (filters, sequential batch runs, unified search, Excel export).
      Stack
      • FastAPI
      • Jinja2
      • Tabulator
    • What I did
      Owned the DB schema, backend integration and frontend wiring of the Java Spring prototype web app. Moved it, with a partner's models 1–2, to company servers and set up its Docker runtime.
      Worked with
      Built together with Vision Inside's in-house Java team; models 1–2 came from a partner company.
      Stack
      • Java
      • Spring
      • MySQL
      • Docker
  6. 06Users
    • Others
      KISTI commissioned the project, and KISTI researchers designed it.
      What I did
      Presented results in regular technical meetings with KISTI researchers and handled data and code requests directly. Wrote the audit deliverables (operator manual, initial data-build report).
ResultThe project led to a follow-on main contractPresented results at KISTI technical meetings · wrote the audit deliverables
Apr 2025 – nowKyutech lab · research assistant (part-time)Joint research with Hitachi

Image-data pipeline · joint research with Hitachi

Rewrote all the software from scratchDesign handed over to two junior researchers

The training-data code I inherited could not be maintained, and every dataset needed a developer sitting next to it. I rewrote everything from camera control to dataset preprocessing as one Python pipeline, with a GUI that non-developer researchers use themselves.

  1. 01On site
    • Others
      The hardware in the lab in the Netherlands was built by a junior researcher who took over the design.
    • What I did
      Worked directly with the camera maker's Spinnaker SDK (PySpin) to connect several cameras, and made their settings adjustable in real time from the PyQt GUI while capturing.
      Equipment
      Several industrial machine-vision cameras (USB3)
      Stack
      • Spinnaker SDK
      • PySpin
      • PyQt
    • Others
      Preparing the samples and the reference labels was done by the project partner's researchers.
      Note
      What the samples are and what is measured are confidential to the joint research, so they are not described.
  2. 02Control and capture
    • What I did
      Built a PyQt GUI with which non-developer researchers capture images and create datasets themselves, without a development environment, and shipped it as a PyInstaller executable. Camera settings are adjusted live from it, and every step runs from this one GUI.
      Stack
      • PyQt
      • PyInstaller
    • What I did
      Wrote from scratch the module that controls several machine-vision cameras, matches their capture settings and saves the images.
      Stack
      • Python
      • PySpin
      • Spinnaker SDK
  3. 03Measurement and analysis
    • What I did
      A program that extracts values from the captures and analyzes them. On site, this program produced the AI training data.
      Stack
      • Python
  4. 04Preprocessing
    • What I did
      Cleaned the captured images and shaped them for training. I replaced the old code stage by stage, checking that the numbers matched on the same inputs.
      Stack
      • Python
  5. 05Output
    • Contents
      An AI training dataset of sample images and reference labels.
      Note
      The research topic and what is measured are confidential to the joint research and are not disclosed.
ImprovedOnly the first frame of dark, low-light video, with a home-made C++ camera layer → single high-quality still captures at the right brightness, through the recommended SDK (PySpin)
ResultDeployed in a lab in the Netherlands, Sep–Oct 2026 · used on site by Hitachi staffDemonstrated to the project partner's researchers · working languages Japanese and English
Aug 2022 – Sep 2023Vision InsideInternal project · international collaboration

Genomic variant annotation and risk report system

Implemented the pipeline, backend and frontendAlgorithm designed with the team lead

A system that links a person's genomic variants (VCF, chip arrays) to public references such as dbSNP, 1000 Genomes, the GWAS Catalog and UniProt, and reports disease and trait risk, population frequencies, and related drugs and herbal medicines. It was an international collaboration with Kyung Hee University's College of Korean Medicine and a Canadian bio research lab. My part implementing the pipeline, backend and frontend, and Docker deployment

  1. 01Input
    • Input
      Raw sequencing data (FASTQ), the input to the variant-calling pipeline.
    • Input
      A VCF file or a genotyping-array (chip) result file uploaded by the customer.
    • Data
      dbSNP (GRCh37, GRCh38), 1000 Genomes, the GWAS Catalog and UniProt.
  2. 02Processing
    • What I did
      Implemented the FASTQ → BWA → samtools → Picard → GATK → VCF pipeline.
      Others
      The pipeline algorithm was designed at the company together with my team lead.
      Stack
      • BWA
      • samtools
      • Picard
      • GATK
    • What I did
      Read customer VCF and chip (array) input with PyVCF and shaped it into the input table used for annotation. Chip input finds its position key in dbSNP by rsID.
      Stack
      • PyVCF
      • Python
  3. 03Storage
    • What I did
      Built dbSNP, 1000 Genomes, the GWAS Catalog and UniProt into MySQL. Split the files by chromosome (sed, split) and loaded them in pandas chunks through SQLAlchemy.
      Stack
      • MySQL
      • pandas
      • SQLAlchemy
      • sed · split
  4. 04Annotation
    • What I did
      A four-stage SQL join: dbSNP and 1000 Genomes on a key of chromosome, position, REF and ALT, the GWAS Catalog on rsID and risk allele, and UniProt on variant ID and allele. A rule-based risk score from GWAS odds ratios and betas, allele frequencies compared across the 26 populations of 1000 Genomes, and drug and herbal-medicine information per disease (MedlinePlus, MeSH, ATC, PharmDB-K, OASIS).
      Note
      The LDA and 1D-CNN models linking diseases to herbal medicines stayed prototype experiments.
      Stack
      • SQL
      • MySQL
      • Python
  5. 05Serving
    • Others
      The screens were designed by a designer at the Canadian bio research lab.
      What I did
      Worked out the requirements with the designer and handled technical communication between the Korean and Canadian teams.
    • What I did
      Served the reports through a FastAPI JSON API and also built the user-facing screens (frontend). Later I removed the web screens from the service and turned it into a JSON API for the frontend.
      Stack
      • FastAPI
      • JSON API
    • What I did
      Set up a Harbor private registry myself on an Ubuntu production server at Korea IDC (certificate, domain, Let's Encrypt) and deployed with Docker images. The production server was kept reachable from internal IPs only.
      Stack
      • Docker
      • Harbor
      • Ubuntu
      • Let's Encrypt
ImprovedWeb screens rendered inside the server → a separate JSON API for the frontend
ResultDeployed to a production server and run for internal IPs only (not a public service)The service code was cleaned up and published
Aug 2023 – Feb 2024Side project · team of two

WJMax game server · play-log collection API

Server, database and securityGame client built by the other member

A server that receives and stores the play reports of a community rhythm game. ~300 concurrent players on average, and up to ~2,000 during updates.

  1. 01Client
    • Others
      The game client was built by the other member of the team.
      Scale
      ~300 concurrent players on average, up to ~2,000 during updates.
      Reports
      Sends the play count of each of a song's eight charts to the server.
  2. 02Edge
    • What I did
      Set up a Cloudflare proxy and WAF in front of the server, so abuse of the unauthenticated reports is stopped before it reaches the server.
      Stack
      • Cloudflare
      • WAF
  3. 03API server
    • What I did
      Ran the FastAPI play-log API in a Docker container on EC2. The public version deals with connection-pool exhaustion, duplicate storage of retried reports, and input checks.
      Stack
      • FastAPI
      • Docker
      • AWS EC2
  4. 04Storage
    • What I did
      Stored the play reports in RDS MySQL. The public version compares song titles exactly as written, so songs that differ only in case or in hiragana vs katakana stay separate.
      Stack
      • MySQL
      • AWS RDS
ResultRan through update days with up to ~2,000 concurrent playersThe public repository is a cleaned-up version (with tests)
02 Research · Kyushu Institute of Technology, graduate school (advisor: Prof. Min-Chul Lee)

3D reconstruction with geometric optics, without training

Research on recovering 3D scenes from images taken with a lens array (integral imaging) camera. It is laid out the same way, as a flow from the input to the result metrics.

ICMV 2026 Best Poster AwardResearch page
  1. Data
    Evaluated on 7 Stanford Lytro light field scenes. The elemental image each lens forms is only a few pixels wide.
  2. Method
    A one-to-one rearrangement that swaps (lens index, pixel position under the lens) (Ray-based Inverse Mapping). Under telecentric capture and orthographic recapture there is no interpolation, so the quantization error is zero.
    Presented
    ICMV 2026, Budapest · Best Poster Award
    More
    RIM and iRIM page
  3. Method
    The rearranged views are used as the first stage of computational integral imaging reconstruction (CIIR) and compared with using the elemental images as they are.
  4. Result
    On the 7 Lytro scenes: PSNR 23.31 → 25.70 dB, SSIM 0.741 → 0.877, and pretrained YOLO detection mAP50 0.665 → 0.776.
  1. Data
    9×9 elemental images, the camera parameters and the depth to reconstruct.
  2. Method
    The search that reconstructed again and again to find the origin with the least quantization error became a calculation from the camera parameters alone. In the paper's experiments it gave the same origin and SSIM as the iterative search.
    Presented
    ITC-CSCC 2025 (Seoul, poster) · ITC-CSCC 2026 (Bangkok, oral)
  3. Result
    For 9×9 elemental images, the reconstructions needed to choose the origin dropped from 81 to 1.
    Stack
    • Python
    • NumPy
    • CuPy
  4. Result
    CPU 0.31 s (measured in the paper), GPU (CuPy, RTX 3090) 0.0057 s. The 25.45 s of the iterative search is an estimate, 0.3142 s per run × 81.
    Code
    origin-point-optimization-vcr
  1. Data
    Views rendered from several positions.
  2. Method
    Because RIM is one to one, it also runs backwards: iRIM synthesizes the elemental image of the virtual lens array you want from multi-view images.
    Presented
    ICCAS 2026, Sapporo · poster
  3. Result
    A 6×6 configuration matched to a 24-inch FHD panel (320×180 lenses). Every view has 100% effective coverage, against 6.7% for the corner views of a simulated lens array camera.
ImprovedReconstructions needed to pick the origin: 81 → 10.31 s measured in the paper · the iterative 25.45 s is an estimate
Lens array elemental image synthesized with iRIM: a torus and a monkey head model split into small lens images
Try it · lens array lab

Capture with a lens array, rearrange and reconstruct, in your browser

Capture a 3D scene through a lens array, rearrange it with RIM, then refocus at each depth. Real Lytro camera data is included too. The lab plays sound.

Open the lab

4 international conference papers

All first author
  1. ICMV 2026Budapest

    Angular-to-Spatial Information Redistribution via Ray-based Inverse Mapping

    Hyeongju Lee, Myungjin Cho, Min-Chul Lee · poster Best Poster Award
  2. ICCAS 2026Sapporo

    Virtual Lens Array Synthesis from Multi-View Images via iRIM

    Hyeongju Lee, Myungjin Cho, Min-Chul Lee · poster
  3. ITC-CSCC 2026Bangkok

    Accelerated Origin Point Optimization for Volumetric Computational Reconstruction

    Hyeongju Lee, Myungjin Cho, Min-Chul Lee · oral · code
  4. ITC-CSCC 2025Seoul

    Image Enhancement of Volumetric Computational Reconstruction Using Origin Point Optimization

    Hyeongju Lee, Gilsu Yeo, Myungjin Cho, Min-Chul Lee · poster · code

Starting point in vision and ML · Depth estimation with correlation filters and VCR, a course project, is where my current research started. Alongside it: an MNIST CNN in MATLAB with backpropagation and momentum gradient descent written by hand, no deep-learning framework. My interest in vision began with my graduation project, Smart Mirror (team of two: Raspberry Pi, OpenCV LBPH face recognition, Kakao speech API). Research path

03 Other work

Other things I built

All work
ProjectWhat I didStack
3DOS · vision enhancement on an XR headsetNov 2024 · XR Device Content Makerthon finalsIn team 3DOS (five lab members registered), I handled development, the presentation and the stage demo. I ported the lab's MATLAB code for Fourier peplography (fog removal) and photon counting (low-light recovery), from Prof. Myungjin Cho's research, to C++ and OpenCV (one teammate helped with part of the port), built them as an Android native plugin (arm64) and demonstrated them in a Unity app on the Metalens 2 headset. The slides are public.
  • C++
  • OpenCV
  • Unity
  • Android
04 History

Experience · education · recognition

CV page

History

Most recent first
  1. Apr 2025 – now
    Kyushu Institute of Technology, Graduate School of Computer Science and Systems EngineeringMaster's program in Creative Informatics · 3D vision research (advisor: Prof. Min-Chul Lee)Finishing my master's in March 2027
  2. Apr 2025 – now
    Kyutech lab · research assistant (part-time)Image-data pipeline for joint research with Hitachi · deployed on site in the Netherlands, Sep–Oct 2026
  3. Aug 2024 – Mar 2025
    Hankyong National University, 3D Image Processing Systems Lab · undergraduate researcherFirst-author poster at Photonics Conference 2024 (Optical Society of Korea) · 3DOS makerthon
  4. Jul 2022 – Nov 2024
    Vision Inside · assistant manager (data and backend engineer)KISTI project data layer · genomic report systemLife-insurance disclosure tool (team of three): loaded insurance disclosure workbooks into MySQL and served them through a FastAPI web viewerCompleted military service as Industrial Technical Personnel (Jul 2022 – Jul 2024)
  5. Mar 2018 – Feb 2025
    Hankyong National University, B.S. in Electrical, Electronic and Control Engineering

Recognition · certification · languages

  • ICMV 2026 Best Poster AwardBudapest · first author
  • 4 first-author international conference papersICMV 2026 · ICCAS 2026 · ITC-CSCC 2026 · ITC-CSCC 2025
  • Rotary Yoneyama Memorial Foundation Scholar2026 · selected through a Japanese-language interview
  • Military service completedIndustrial Technical Personnel (Jul 2022 – Jul 2024)
  • SQLDOct 2023 · Korea Data Agency
  • LanguagesKorean (native) · Japanese (business-conversational) · English (conversational)
05 Contact

For roles and collaboration, please write by email.

I am looking for roles in data engineering, backend development and vision data pipelines.

Finishing my master's in March 2027