Data 2025–2026 · Ongoing · Client: Winkoms (in-house product) · Responsible: Francesc Xavier Gómez
The challenge
Every applied computer-vision project starts with the same bottleneck: turning raw microscopy images into labeled training data, then training and running models against them — usually across separate, disconnected tools with no shared access control, GPU management, or usage tracking.
The solution
Winkoms built WNK Platform, a single authenticated web platform that hosts three integrated services behind one login:
- YOLO Labeler — image labeling for YOLO training, with YOLO + SAM 2.1 auto-segmentation to speed up annotation
- QaiFIL Analysis — a Streamlit-based filament detection and quantification tool (the same engine behind the QaiFIL fibre-optic sensing project for UB)
- Trainer App — GPU-backed YOLO model training, limited to one concurrent training job to protect shared GPU resources
The platform runs on a dedicated NVIDIA GPU 24 GB VRAM and tracks usage per service and per user — images labeled, analyses run, training jobs launched — through a lightweight admin panel. The interface is currently in Catalan, and translates cleanly to English or Spanish with browser tools like Google Translate if needed.
The results
A single, multi-user platform now unifies three previously separate AI tools under one authenticated entry point, with live GPU monitoring and per-service usage accounting built in from day one.