2025년 2학기 Gradi 개발 완료.

[Downy-newlearner]

FE & BE
어플리케이션 / 웹 / 디바이스 중 어플리케이션을 일부 완성했다.
정상적으로 구동되고, AI 서버와도 잘 연결되었지만 실사용을 고려했을 때는 부족한게 매우 많다.

AI
2026 수능완성 영어 문제집 한 권, 290문제에 대해 99%의 채점 완수율을 보였다.
하지만, Detection(YOLO), Recognition(ResNet) 모두 Seen data에 대한 추론이었기에, 99%의 완수율은 큰 의미가 없다.
또한 객관식에 대한 채점이었으므로, AI에서의 할 일은 너무나 많이 남아있다.

오는 2026년에는 다음 작업을 우선으로 해야한다.

• 다양한 문제집 데이터(2만장 이상, 60권 분량 이상)를 모두 학습시킨 Detection Model 만들기
• 소형 학원을 타겟으로 하여 해당 학원들이 취급하는 문제집에 대해 Textbook Identification 만들기(DB 만들기)
• OCR 등 Recognition Solution 제시하기

우리 팀원들 너무 고생했고, 좋은 연말 보냈으면 좋겠다.

---

Note

Adds a README summarizing answer-classifier experiments and a MNIST base-model training notebook with logs/plots.

• Test/Classification:
  • Docs: Add answer_classifier_README.md detailing dataset, training setup, and CV results for EfficientNet-B0, ResNet18, and MobileNetV3 answer classifiers (answer_1/answer_2).
  • Notebook: Add mnist_base_model.ipynb implementing and logging MNIST training runs (EfficientNet-B0, ResNet18), including metrics, model saving, and plot outputs.

^{Written by Cursor Bugbot for commit 064a6c2. This will update automatically on new commits. Configure here.}

[cursor]

This is the final PR Bugbot will review for you during this billing cycle

Your free Bugbot reviews will reset on January 7

Details

You are on the Bugbot Free tier. On this plan, Bugbot will review limited PRs each billing cycle.

To receive Bugbot reviews on all of your PRs, visit the Cursor dashboard to activate Pro and start your 14-day free trial.

[cursor]

Bug: Missing output_json_path attribute initialization causes failure

The LLMDBPipeline class uses self.output_json_path in _init_json_file and save_to_json methods, but this attribute is never initialized in __init__. When the constructor calls _init_json_file(), it will raise an AttributeError (caught by try-except, logging an error). The JSON initialization silently fails, and any subsequent calls to save_to_json will also fail. The groq_api_key parameter is also accepted but never used, suggesting incomplete implementation.

Additional Locations (2)

• ai-service/DB/files/crop_images.py#L43-L51
• ai-service/DB/files/crop_images.py#L52-L69

 

[cursor]

Bug: All problems in section get same content detections

In crop_and_save_images, when iterating over multiple problems within a section, the code always assigns the same content to every problem. For english_content, korean_content, and answer options, it filters all detections of each type from the section then uses eng_dets[0], kor_dets[0], etc. for every problem. This means if a section has 3 problems, all 3 get saved with identical english/korean content images (just with different filenames like prob_00_english_content.jpg, prob_01_english_content.jpg). The content detections need to be matched to their corresponding problems, likely by spatial proximity to each prob_det.