PROJECTS · REVIEWS · NOTES · OCCASIONAL BAD IDEAS
Radiotherapy software. Open models. iPhone apps. Business ideas. Hardware. Competitions. Things I tried that did not work. Whatever seems worth looking at.
Read the story →
Lunit Medical Foundation Model HackathonBenchmark Award · Team AIM · 1st place
Result, method & limits →
Loose notes
Some things worked.
Some really didn’t.
The code, screenshots, calculations, build files, caveats, and abandoned assumptions can all live here.
Start here
Three ways into the work—whether you have a problematic DICOM dataset, need an implementation note, or want to operate a medical challenge.
Inventory objects and flag obvious privacy or reference problems without uploading patient files.
Open the inspector → PRACTICAL CASE STUDYSeven final plans, 36 structures per plan, hidden scoring, and an anonymized 150-point leaderboard.
Read the operating note → A PRODUCT QUESTIONAn experiment in giving societies, hospitals, and research groups the infrastructure to operate their own reproducible competitions.
Read the hypothesis →From the build log
Field Notes · Secure evaluation
I built the first local Orcaton evaluator for submitted AI containers, then wrote down both the isolation policy and the Docker failure that stopped the first end-to-end run.
Read the field note →Three threads
Not a feed of everything I do. Three ongoing bodies of work, each grounded in experiments and evidence.
Clinical systems
Radiotherapy, dose, DICOM, treatment planning, QA, and the gap between research software and the clinic.
Explore the field notes → 02Code & prototypes
From a rough question to code, tests, CAD, BOMs, and a prototype someone else can inspect.
See how things get built → 03Ideas under pressure
Product ideas, market reality checks, failed assumptions, pricing tests, and explicit reasons to stop.
Read the experiments →Selected work
Clinical software, browser tools, physical builds, and speculative infrastructure carried through to working systems.
Executable science fiction · Distributed systems
Mission control for a star-sized distributed system. A deterministic browser simulation lets an operator stress 10,000 autonomous solar collectors with communication loss, thermal limits, debris, demand spikes, and cascading failures.
Radiotherapy · Patient tools
A stateless web service that turns a patient's DICOM-RT records into CT dose overlays, structures, DVHs, treatment summaries, and plain-language explanations.
Local vision · Resale
SellShot and PackCheck analyze product-photo quality and packaging changes entirely inside the browser, without uploading images to a server.
Hardware · Resale
A repeatable eight-angle photography turntable for secondhand products, including CAD, STL files, ESP32 firmware, iPhone control, and a complete BOM.
Plant tech · Fabrication
A propagation cabinet that records rooting and acclimation. The HOME 12 and PRO 48 studies include build guides, costing, CAD, and explicit safety boundaries.
Treatment planning · Monte Carlo
An open, reproducible radiotherapy planning stack spanning DICOM import, fluence optimization, MCO, MLC sequencing, virtual delivery, and independent photon, proton, and carbon Monte Carlo experiments.
Claude · Clinician-in-the-loop
A clinician-supervised planning workspace where an agent proposes, optimizes, compares, and revises research plans while every consequential step remains subject to human approval.
Radiotherapy · Benchmarking platform
Built and operated for a real treatment-planning competition: seven final plans were evaluated across 36 structures each and ranked on an anonymized 150-point leaderboard.
Medical AI · Reproducible challenges
I kept wondering why a medical society should need an equipment vendor—or its own software team—to run a credible challenge. Orcaton is a working experiment in another model.
Agent skills · Active recall
AI wrote the code. Now make it yours. A Claude Code and Codex skill that turns a real Git diff into one-question-at-a-time lessons and code-grounded project interview practice.
Local AI · Smart glasses · AWS Strands
I built a consent-first smart-glasses agent that interprets selected moments on the phone, minimizes what leaves the device, and proposes one explainable next action without silently taking it.
iPhone apps
SwiftUI applications built around cameras, QR codes, Bluetooth devices, labels, and local records.
SwiftUI · Vision OCR · Camera
A local-first move-in and move-out evidence app. It reads room names from a floor plan, guides both parties through the same nine views, compares matched captures, records signatures, and exports a PDF report.
SwiftUI · Local-first operations
An iPhone inventory, QR checkout, and return-inspection app for small rental operations. It records kit components and condition photos, prints AirPrint or ZPL labels, and stores data locally.
Swift · BLE · Camera
An iPhone controller that coordinates an ESP32 turntable, lighting, homing, weight readout, and an eight-angle product photography sequence.
Medical AI research
Public research notes on medical-image restoration, segmentation, radiotherapy, and DICOM.
Failure, documented
15confirmed rejectionsVUNO twice. Yonsei Medical Center after three separate interview processes. Then Lunit, Palantir, Accuray, Kakao Healthcare, Radformation, and Elekta in one 2026 run. This is the unpolished record behind the projects above.
Read the rejection log →About
My background is in medical physics and medical AI. I now explore local AI and buildable product systems alongside clinical software. For research collaboration, technical consulting, or prototype review, reach me through GitHub, LinkedIn, or Kmong.