Responsibilities
- Design, train, debug, and improve state-of-the-art object detection models for real-world conditions
- Build robust training pipelines: datasets, augmentation, caching, versioning, and reproducible experiments
- Perform systematic error analysis and ablations to isolate failure modes (data vs model vs inference vs post-processing)
- Develop custom detection systems beyond standard training, including multi-stage pipelines, ensembles, and specialized post-processing
- Optimize inference for latency, throughput, and memory, including GPU acceleration and export toolchains
- Deliver production-grade services using Docker, Linux, CI/CD, and APIs (FastAPI and/or gRPC)
- Implement testing strategy across the pipeline (unit, integration, regression), including “golden image” test sets
- Set up monitoring and maintenance: logging, metrics dashboards, drift/performance tracking, retraining triggers
- Write clear technical documentation, architecture decisions, and trade-off analyses
- Read research papers and rapidly translate ideas into working prototypes and deployable components
Requirements
- Python and ML Engineering
- Advanced Python engineering: clean architecture, packaging, typing, testing, profiling
- Strong PyTorch experience (must)
- Strong model debugging skills and disciplined experimentation
- Experiment tracking and reproducibility: W&B and/or MLflow, deterministic runs, seed control
- Config management: Hydra and/or OmegaConf
- Data pipelines: PyTorch Dataset/DataLoader, augmentation pipelines, caching
- Dataset versioning: DVC or equivalent
- Computer Vision Fundamentals
- Strong CV fundamentals: preprocessing, geometry, photometric effects, distortions, camera models
- OpenCV expertise for classical CV and integration into modern ML pipelines
- Evaluation expertise: mAP, precision/recall, IoU, PR curves, calibration
- Deep Object Detection Expertise
- Hands-on experience with modern detectors such as: YOLO (v5/v8/v9), Faster R-CNN, RetinaNet, EfficientDet, DETR variants
- Experience building advanced detection workflows: Multi-stage detection (proposal, refine, classify)
- Ensemble and stacking strategies
- Specialized post-processing tuned to domain constraints
- Production ML and MLOps Delivery
- Model export and serving: ONNX export/runtime, plus at least one of TorchScript or TensorRT
- GPU inference optimization and performance tuning (batching, throughput, latency, memory)
- Deployment: Docker, Linux, CI/CD basics (GitHub Actions and/or GitLab CI)
- Service implementation: FastAPI and/or gRPC, model versioning, rollback strategy
- Monitoring and lifecycle: drift/performance tracking, logging, dashboards, retraining triggers
- Testing: unit tests for preprocessing/post-processing, integration tests, regression sets, threshold stability tests
- R&D Capability
- Ability to read papers and implement ideas quickly
- Strong debugging methodology, ablation design, and error analysis
- Clear technical writing and engineering decision-making
Nice to Have
- Engineering Drawings
- Domain Experience with engineering drawings and technical documents
- PDF vector vs raster workflows, line detection, symbol detection
- Table/diagram understanding, CAD-like concepts, annotation workflows
- OCR + vision hybrid systems (even if not OCR-first)
- Document and Diagram Vision Toolchain
- PyMuPDF and/or pdfplumber
- Image rasterization, coordinate transforms
- Handling noisy scans: skew/warp correction, deskewing
- Broader CV Capabilities
- Instance segmentation: Mask R-CNN, YOLO-seg
- Keypoints, pose, landmark detection
- Tracking for video: ByteTrack, DeepSORT
Additional Information
- Working Days: Sunday to Thursday. (Fri & Sat remains off)
