Geonhee Kim

Web development · Computer vision · Operations automation

geonhee@forblune.com · github.com/forblune · forblune.github.io

Summary

I write down whether what I built is up right now, and what it is not, in the same table. I train computer vision models and ship them as far as an app, I build and run responsive sites, and I find the places where something already running broke quietly with evidence anyone can reproduce. Every number below is one I measured myself. I do not present self-initiated work as paid client work.

Selected projects

RePET Vision — automatic sorting of clear PET bottles · Sep 2025 · Gangwon AI Idea Competition Two-class object detection, clear and coloured PET. 700 images collected through Roboflow, augmented to 1,726 (train 1,539 / val 126 / test 61), YOLOv8n · 640px · 40 epochs · batch 16. mAP50 approx. 91% · mAP50-95 approx. 52%. Built through to an Android Studio app with live camera recognition and confidence display. Attached an economic case using Gangneung city recycling rates, 330 → 538 KRW/kg.

K2J — smoking detection on CCTV · Jul 2025 · CKU summer AI project Single-class detection on park CCTV footage. YOLOv5 · PyTorch · 640px · 50 epochs. mAP@0.5 approx. 0.97 · mAP@0.5:0.95 approx. 0.68 · Precision approx. 0.90 · Recall approx. 0.98. I owned the progress record, wrote the video boundary-correction code, and built the frame-interval feature.

MindHub — mental health care platform · Jun 2026 to present Patient daily records → AI clinical summary → handoff to the clinician, and CBT homework prescribed back the other way. React · Supabase · Kakao login · Solar (Upstage) LLM. Scoped deliberately as recording, summarising and handoff, not diagnosis.

Eight live sites — addresses and boundaries are tabulated at forblune.github.io. Web Care (pricing, refunds, enquiry, checkout), Style Gallery (8 directions, 24 pages, 8 in English, single-file HTML with no framework), Mellow Room · OpsFlow · ClientFlow · ServiceOS (self-initiated demos, fictional data). All of them respond right now.

Problems solved

Cases from sites that are live. One of the four below is an incident I caused, and I wrote that into the document as it happened.

Enquiries were not arriving and the page looked fine — a static-only deploy dropped the server function, and the enquiry API returned 405 for about four hours. Separating 400 (bad request) from 405 (path exists, method gone) settled "since when, and because of what." After recovery I froze deploy verification into a procedure — check the deployment URL, not the domain, because a CDN cache will return 200 for a file that is already deleted.

Nobody knew where the current version was — the source of a live site was in no git repository. Hashing eight local candidates against live gave 0 matches. I fixed the live output as canonical and made the deploy one-directional.

Someone else's outage became our outage — a free external image service was hotlinked directly, so half the page turned into grey boxes during their incident. 243 external hotlinks → 0, with 234 files selected as CC0 and moved to self-hosting.

The real cause of "it looks AI-generated" — usually the fundamentals rather than the content. I quantified Korean breaking mid-word using Intl.Segmenter and Range.getClientRects(), applied it across every live site and 25 example pages, and brought it to 0.

Awards

Business and operations

Open-market advertising operations · Jan–Mar 2025 — managed ad groups for around 1,700 products across Gmarket, Auction and 11st. Improved ROAS through ad on/off decisions, category matching, and keyword and daily-budget adjustment, and made the calls on restructuring the catalogue.

Naver Smart Store, sole operator · Oct 2024 to present — ran the full cycle under my own business registration: sourcing, listing, orders, shipping, exchanges and customer service, and settlement.

Other — NYPC Code Battle entrant (Jul 2025, competitive algorithm strategy) · RowCraft (May 2026 to present, a solo data-cleaning business on Fiverr, with an SOP built on pandas and rapidfuzz plus a Slack bot) · BTCUSDT paper-trading bot running continuously on a Raspberry Pi · water purifier lever designed and 3D printed from scratch · Raspberry Pi 5 and Arduino sensor work.

How I work

I do not fix by guessing. I build a reproducible boundary that settles the cause first. Code compiling and code working for a user are different things, so I confirm the fix on the real screen. I leave why it happened and how to check it in the repository, so the same incident does not happen twice. I write down my own mistakes too — hiding them costs the credibility of everything else I report.

Education and skills