DIATICS · May – Sep 2024
Full-cycle computer vision
- Role
- AI Intern — full cycle from data preparation and labeling through training, evaluation, and serving
- Dates
- May – Sep 2024 · Istanbul
- Stack
- YOLOv8FastAPIPostgreSQL
- fire-detection false alarms
- −30%
- plate recognition accuracy
- 92% → 97%
- pose estimation precision
- 89%
Deep learning for computer vision — object detection, classification, and pose estimation.
Three workstreams
Fire detection
A YOLOv8 model for detecting fire. False alarms dropped 30% and response improved 45%.
Plate recognition
Recognition accuracy went from 92% to 97%, and identification speed improved 40%.
Pose estimation
89% precision, with processing time down 35%.
This was intern work inside a company, so I keep the detail here to what is on my CV: the outcome of each workstream, with the specifics belonging to the employer.
Serving
The models did not stop at evaluation: I served them through FastAPI backed by PostgreSQL, and retrieval time dropped 60%.