Role & responsibilities

About the Role

We are hiring an AI Developer focused on Computer Vision to build and ship vision models that run in production. This is a hands-on engineering role you will spend your time training detection models, cleaning and labeling data, squeezing latency out of inference pipelines, and getting models off your laptop and onto real hardware.

You will work closely with product and backend engineers on problems like [object detection / OCR & document extraction / video analytics / defect inspection / people counting / ANPR].

Key Responsibilities

Model Development

- Train, fine-tune, and evaluate object detection and segmentation models — YOLO (v5/v8/v11), Detectron2, RT-DETR, or equivalent.
- Run structured experiments: hyperparameter tuning, augmentation strategies, anchor/loss configuration, and ablation studies.
- Diagnose model failures on real-world data — occlusion, motion blur, lighting variation, small objects, class imbalance — and fix them at the data or architecture level.
- Report results with proper metrics (mAP@50, mAP@50-95, precision/recall, IoU, confusion matrices) rather than eyeballed demos.

Data Pipeline

- Build and maintain data pipelines: collection, cleaning, annotation QA, dataset versioning, and train/val/test splits that don't leak.
- Manage annotation workflows using tools such as CVAT, Roboflow, or Label Studio; write scripts to convert between COCO, YOLO, and Pascal VOC formats.
- Generate synthetic or augmented data where real data is scarce.

Deployment & Optimization

- Convert and optimize models for inference — ONNX, TensorRT, OpenVINO, quantization (INT8/FP16), and pruning.
- Build inference services with FastAPI/Flask, containerize with Docker, and deploy to [cloud / on-prem / edge devices].
- Work with real-time video pipelines: RTSP streams, OpenCV, GStreamer, frame batching,



and multi-object tracking (ByteTrack, DeepSORT, BoT-SORT).
- Profile and optimize for FPS, GPU memory, and end-to-end latency.

Engineering Practice

- Write clean, tested, reviewable Python; use Git properly.
- Track experiments with MLflow, Weights & Biases, or similar.
- Monitor deployed models for drift and degradation; support retraining cycles.
- Document what you build so the next person can run it.

Required Qualifications

- 2–4 years of hands-on experience building computer vision applications.
- Bachelor's in Computer Science, Electronics, Engineering, or equivalent practical experience.
- Robust Python, plus NumPy and OpenCV.
- Practical experience with PyTorch (TensorFlow also acceptable).
- Demonstrated experience training and deploying YOLO-family models — you should be able to explain what you tuned, why, and what it did to your metrics.
- Solid grasp of CV fundamentals: CNNs, convolution/pooling, transfer learning, NMS, IoU, anchor boxes, and the precision/recall trade-off.
- Experience taking at least one vision model from dataset to a working deployed service.
- Comfortable with Linux, Git, and Docker.

Preferred Qualifications

- Experience with edge deployment (NVIDIA Jetson, Raspberry Pi, Coral, or similar).
- Multi-object tracking or pose estimation in production.
- Familiarity with vision transformers (ViT, DINOv2) or open-vocabulary detection (SAM, Grounding DINO, YOLO-World).
- OCR experience (PaddleOCR, Tesseract, TrOCR) or document AI.
- Exposure to cloud ML services (AWS SageMaker / EC2 GPU, GCP Vertex AI, or Azure ML).
- 3D vision, depth cameras, stereo, or point clouds.
- Basic C++ for performance-critical inference paths.
- Kaggle results, GitHub projects, or open-source contributions in CV.
- Domain exposure in [manufacturing / retail / security & surveillance / healthcare / agriculture / logistics].

Preferred candidate profile

📌 AI Developer (Nagpur)
🏢 HSM EDIFICE Construction Services
📍 Nagpur