Chapter 02 of 06 · Product Engineer arc
02 プロジェクト[redacted]
Project [redacted]
A private Google Photos, with its own machine learning!!
↓ Google Photos–style platform with custom ML. Sprout Digital Labs · client project.
- Role
- Lead on 3 of 4 detection pipelines
- Reach
- Government client · NDA
- Period
- Nov 2024 – Sep 2025
Under NDA. Happy to walk through the engineering in an interview.
活躍What I did
-
Built the age/gender estimation, landmark recognition, and notable-figure recognition pipelines of a computer-vision microservice (FastAPI, Docker), inside a service a colleague had scaffolded.
-
Ported an earlier Flask prototype into the production service, replacing stubbed endpoints with real model inference.
-
Integrated an open-source age/gender model (MiVOLO) with a YOLO-based face and person detector.
-
Replaced a TensorFlow/Keras landmark classifier with a lighter YOLO + ONNX pipeline, removing the TensorFlow dependency entirely.
-
Fixed model loading after a PyTorch 2.6 upgrade changed a default and broke every model.
-
Added GPU detection and traceback-level error logging for production debugging, plus container and ARM64 build fixes.
-
Wrote a retraining guide for the two custom-trained models.
-
Extended the analysis results page with similar-photos and detail views, in English and Indonesian.
-
Reworked global search into people, places, and folders with debouncing and caching, and fixed a regression that had switched off semantic search ranking.
-
Made the map auto-center and zoom to fit photo locations.
-
Removed an overly strict request-size limit in a production hotfix.
成果 Impact!!
教訓 What I learned
