Multi-Class Object Detection System

I’m looking for an engineer who can take full ownership of a computer-vision pipeline that reliably spots vehicles, persons, and animals in images or short video clips. The goal is to move from raw data to a production-ready detector exposed through a lightweight FastAPI endpoint that I can drop into my existing stack.

Here is what I need done:
• Curate or expand an annotated dataset that covers cars, trucks, bikes, pedestrians, and common domestic or wild animals in varied lighting and weather conditions.
• Train and fine-tune an object-detection model—YOLOv8, Faster-RCNN, or another state-of-the-art architecture in TensorFlow or PyTorch is fine as long as it balances accuracy and inference speed.
• Validate the model with mAP, precision/recall, FPS, and confusion-matrix reports, then iterate until it meets the agreed thresholds.
• Package the final weights and inference code behind a FastAPI REST endpoint that accepts an image file or base64 string and returns JSON bounding boxes, labels, and confidence scores.
• Provide a brief Dockerfile so I can deploy the service to an Azure Container Instance; GPU acceleration should be auto-detected when available.
• Hand over clean, commented source code, the trained weights, reproducible training scripts, and a short README explaining setup, retraining, and extension steps.

Acceptance criteria
1. mAP@0.5 ≥ 0.85 on a withheld test set for each of the three classes.
2. End-to-end latency (image upload → JSON response) < 350 ms on an NVIDIA T4 GPU.
3. Repository passes a fresh clone test: `docker compose up` spins up the API and a sample request returns detections.

If this sounds like a challenge you’re eager to tackle, tell me briefly which model you’d start with and why.