Role Summary
AI/ML Developers / Data Scientists own the quality of the models behind the platform’s drone video and image analytics. You will improve detection, segmentation and tracking accuracy on aerial imagery, adapt models to each customer’s domain, and build the benchmarks and monitoring that keep results trustworthy in production.
Key Responsibilities
Modelling
- Detection and segmentation: Train and improve object detection and instance/semantic segmentation models for drone footage (e.g. YOLO family, RT-DETR, Faster/Mask R-CNN, U-Net, SAM-assisted labelling).
- Aerial-specific challenges: Address small and dense objects, oblique angles, scale variation with altitude, motion blur and lighting using tiling/slicing inference (SAHI), multi-scale training, mosaic and photometric augmentation, and altitude-aware sampling.
- Tracking: Implement multi-object tracking (ByteTrack, BoT-SORT) with camera-motion compensation; produce persistent object IDs across frames and flights.
- Geo-referencing: Work with the Drone Data Architect to convert pixel detections to geographic coordinates using camera pose and terrain data, and quantify positional error.
Data and Fine-tuning
- Dataset curation: Build versioned datasets (DVC or equivalent) with annotation guidelines, labelling QA (inter-annotator agreement), class-balance analysis and leakage-safe splits by site and flight rather than by frame.
- Active learning: Select high-value samples using uncertainty and embedding diversity; run labelling in CVAT or Label Studio.
- Per-customer adaptation: Define a repeatable fine-tuning workflow (transfer learning, freezing strategies, LoRA/adapters where applicable) with hard guarantees that one tenant’s data never trains another tenant’s model unless explicitly permitted.
Evaluation and Monitoring
- Benchmarks: Create golden test sets and report mAP@0.5 and mAP@0.5:0.95, per-class precision/recall/F1, confusion matrices, tracking metrics (MOTA, IDF1) and latency/throughput per model version.
- Release gates: Define regression thresholds that block promotion of a model that underperforms the current production version.
- Drift: Monitor input drift (image statistics, embeddings, PSI/KS tests) and output drift (class distribution, confidence); set retraining triggers and alerts.
- Explainability: Provide confidence calibration and failure-mode analysis for product and customers.
Optimisation and Handoff
- Export: Export models to ONNX; support TensorRT optimisation with FP16/INT8 calibration and verify accuracy loss against benchmarks.
- Collaboration: Hand over model artefacts, pre/post-processing specs and test sets to the Python AI/ML developer; validate MVP-to-SaaS result parity for migrating customers.
Required Skills and Experience
- 5+ years in applied machine learning, including 3+ years in computer vision on video or aerial imagery.
- Strong PyTorch and OpenCV; experience training detection and segmentation models end to end.
- Rigorous evaluation methodology: mAP, PR curves, per-class analysis, statistical validation.
- Dataset curation, annotation tooling and augmentation pipelines.
- Python engineering practices: testing, packaging, reproducibility, experiment tracking (MLflow or W&B).
- Working knowledge of ONNX export and GPU training.
Preferred / Added Advantage
- SAHI/tiled inference, SAM or foundation-model-assisted labelling.
- Multi-object tracking and video analytics.
- Drift detection tooling (Evidently or similar).
- TensorRT quantisation; edge deployment (Jetson).
- Vision-language models for search or auto-labelling.
- Remote sensing or geospatial ML.
Key Deliverables
- Baseline and improved models with benchmark reports per release.
- Versioned datasets and labelling guidelines.
- Fine-tuning pipeline and tenant-isolation rules.
- Drift monitoring spec and dashboards.
- Model cards for each production model.
Education
Bachelor’s or Master’s degree in Computer Science, Data Science, Engineering or related field.
Pay: ₹600,000.00 - ₹1,000,000.00 per year
Work Location: In person
