Detection · Segmentation · Tracking · Edge Deployment
Pune, India · shindepranav.site · shinde.a.pranav@gmail.com
I build CV systems end-to-end — from raw data and annotation infrastructure through model training to production inference. Currently at a defence R&D firm working on object detection and tracking for real operational environments.
Things I've shipped in production:
- 500k+ image annotation pipeline — bare-metal CVAT (PostgreSQL/Redis/Nginx), YOLO cascade for initial masks, custom SAM 2.1 boundary refiner, watchdog-driven export→process→re-upload loop. No manual triggering.
- 13-class instance segmentation — mAP50 ~0.74, custom
WeightedDetectionLossfor severe class imbalance, P2 feature pyramid heads for separating visually similar classes in cluttered scenes. - PyTorch → TensorRT INT8 migration — 9.2 GB → 2 GB memory footprint (−78%), ~23 ms → ~3.7 ms latency. Profiled with Nvidia Nsight Systems/Compute across RTX 3060 and RTX 5080.
- Dual-GPU training pipeline — resolved CUDA context conflicts between training and validation processes, cut validation turnaround by 18%.
| Area | Tools |
|---|---|
| Detection & Segmentation | YOLO v8/v11, SAM 2.1, OpenCV, custom loss functions |
| Tracking | ByteTrack, BoT-SORT, Kalman Filters, ReID |
| Optimization | ONNX, TensorRT, INT8/FP16 quantization, Nsight profiling |
| Training Infra | PyTorch, multi-GPU, CVAT REST API automation |
| MLOps & Cloud | Docker, Kubernetes, AWS (EC2/S3/Lambda/RDS), CI/CD |
| Backend | Python, Node.js, PostgreSQL, Redis |
Benchmark your YOLO model across every export format on your own hardware — PyTorch, ONNX, CoreML, TensorRT — and get real FPS, real latency, real accuracy delta. Because "ONNX is faster" depends entirely on your machine.
python onnx tensorrt coreml benchmarking �� 🚧 active development · join waitlist
- 🎓 B.E. in AI & Data Science — MMCOE, Pune University (2023–2026)
- 🏆 Top 10 National Finalist — Smart India Hackathon 2024 (50,000+ competing teams)
- ☁️ AWS Certified Developer Associate (2024)
- 🪟 Microsoft Azure Fundamentals AZ-900
Ship the model, not just train it.