Software engineer (new grad · Seoul · open to remote). I build Unity IL2CPP reverse-engineering support and check the measurements behind pose-estimation tools.

• Wrote the original Unity IL2CPP support in rea (reverse-engineering toolkit, 400k+ stars); it shipped in 6.8.0
• 29 merged PRs in 14 external open-source repos: rea 7 · Ultralytics 4 · SurfSense 3 · supervision 2 · RF-DETR 2 · 11 more in 9 other repos (list). Not counted: my own repos and team project gaeun-k/ADS.
• Wrote the first version of supervision's COCO keypoint mAP metric, which the maintainer extended and shipped in 0.31.0
• Author of poseaudit and Epokio (both on PyPI)

📧 bkbkjang@gmail.com · 🎮 itch.io

rea (reverse-engineering toolkit) · #3807, #5326, #5327, #5785, #6989

• Problem: in a stripped Unity IL2CPP build, method names sit in global-metadata.dat and the code in a separate native binary. In one Android game, Ghidra alone found 128 functions; seeds from outside tools such as Il2CppDumper created the other ~117,000.
• My part: wrote the original IL2CPP implementation:
  • a global-metadata.dat reader (metadata versions 24, 27, 29, 31 and 39)
  • linking methods to native addresses in PE, ELF and Mach-O, with Ghidra seeds so rea function <binary> 'Type$$Method' decompiles by name
  • later, generic method instances (#5785) and arm64 builds inlined by link-time optimization (#6989)

_{Unity 2022.3 sample built for this demo, Ghidra 12.1.4, rea main 1173c0a9d}

• Verified against Il2CppDumper:
  • by two users on commercial builds (binaries not shared): all 111,771 non-generic names on Windows x64; all 58,901 metadata-linked addresses on Android, after a zero-fill fix the tester proposed
  • by me on a public MIT-licensed Unity build (Between Two Worlds 1.0.5): 43,198 Windows and 42,098 Android code rows, 0 differences
• Result: squash-merged by the maintainers (gensokyo-dev, and morluto on #6989), who substantially reworked #5326 and #5327 (bounds checks, pointer-table and candidate handling) and added a compact query API and a performance fix. Released in rea 6.8.0.
• Earlier: paired same-named nested .NET types by their enclosing types (#3079), released in rea 6.6.0.

Ultralytics (YOLO) · #26564, issue #26563

• Problem: convert_coco(use_keypoints=True) kept people with no labelled keypoints, so pose validation counted them as targets and under-reported pose mAP. The official COCO pose numbers were unaffected; it hit data converted with convert_coco and custom labels with fully occluded people.
• My part: found and reproduced it against pycocotools, and proposed a validator-side fix (yolo11n-pose on a 200-image COCO val2017 sample: pose mAP50-95 0.341 → 0.486).
• Result: the maintainer (glenn-jocher) fixed it at the source with a one-line change in the converter instead. With that fix, the maintainer measured a different model, yolo26n-pose, on my 200-image sample going from 0.347 to 0.576, matching pycocotools.

supervision (Roboflow) · #2687

• My part: wrote the initial sv.metrics.KeyPointMeanAveragePrecision (COCO keypoint mAP using OKS) and its first pycocotools parity tests; on 200 COCO val2017 images, AP, AP50, AP75, APm and APl matched pycocotools within 3e-9.
• Result: the maintainer extended it before release. It shipped as one of the headline features of 0.31.0, whose release notes credit me for it and for fixing perfect-prediction mAP scoring.
• Earlier: fixed from_yolo reading pose labels as polygons (#2655), later extended by the maintainer and released in 0.30.8.

poseaudit · Python · PyPI · Browser demo
Checks how far angles and lengths computed from pose keypoints disagree with the labels, the way a measuring device is validated (bias, limits of agreement, ICC). On 322 left elbows from 200 COCO val2017 images, YOLO11n-pose disagrees with the labels by 15° or more on 58% of small arms (under 30 px) vs 26% of large ones (over 60 px). How much of that is label noise is still open; simple noise models do not explain it (details).

_{Photos from COCO val2017, all CC BY-NC-SA 2.0: Hans-Peter Gilgen, Marcelo Costa, Flickr photo 6886950764. Modified with keypoints and angles drawn; this GIF is shared under the same license. COCO annotations CC BY 4.0.}

Epokio · Python · PyPI
Sends a phone alert when an ML training run stalls, hits NaN or finishes, by reading the logs it already writes (Ultralytics, Hugging Face, Lightning, Keras, TensorBoard). No code changes, no account. Listed in ntfy's integrations.

lidtilt · Python Reads the MacBook lid angle sensor to remove the lid's tilt from webcam head pitch.

Stella Ball · browser game A small browser game. Play on itch.io →

• ultralytics#26934: with perspective augmentation, objects at Mosaic canvas corners could cross the projection plane and flip into image-sized boxes. I added a filter that drops them; the maintainer folded it into the existing box filter, which now drops only objects that cross the plane. Released in v8.4.193.

  
  _{Synthetic canvas, perspective=0.001, seed 2. Over seeds 0 to 999 of this setup, 58 frames got a box over a third of the image in 8.4.192 and 0 in 8.4.193.}

• ultralytics#26651: SAM prompts for a folder or video reached only the first image. I made them apply to the whole source. In review the maintainer caught that my version modified the caller's arrays in place; I fixed it, and they moved the forwarding to the streaming entry point.
• ultralytics#26635: every DataLoader worker drew the same Albumentations augmentations. I traced it and proposed reseeding per worker; the maintainer reworked the final commit.
• rf-detr#1653, #1654: loading a pretrained RF-DETR logged warnings that did not apply: DINOv2 weight warnings on 12 of the 14 default configs, and a class-count warning that said the opposite of what happened. I made them match what loading actually does.
• Improve YouTube#4347: a sped-up livestream that caught up to the live edge got stuck stalling and rebuffering. I made it drop back to 1x there, as YouTube's own speed menu does. Released in v4.2400.

All merged pull requests →