AI Engineer — Computer Vision & On-Device Detection (Accident & Driver Monitoring)

Kasava AI Pvt Ltd · Kolkata · Full-time Kasava turns dashcam footage into tamper-proof, court-ready evidence for fleet insurers, automating claims, cutting fraud, and lowering loss ratios. Everything downstream depends on understanding, from raw camera and sensor data, what is happening on the road and inside the cab: the driving context and risk moment-to-moment, the distraction and unsafe-driving behaviours, the accidents and near-misses, and exactly what the conditions were at the moment of impact. We're looking for a senior computer-vision engineer to own that detection layer across both cameras, building the models that understand the road scene and watch the driver, running them efficiently on our own dashcam hardware, and turning every event into a clean, timestamped, context-enriched record that the claims and risk pipeline can trust. WHAT YOU'LL BUILD • Forward-facing road-scene understanding: continuous perception of the driving environment from the road camera — detecting and scoring risk events (tailgating, harsh braking/cornering, near-misses, lane departures, traffic-signal and hazard context, adverse road conditions), with accidents and collisions as the highest-stakes case, fused with telematics (IMU / accelerometer / GPS) and held to the false-positive discipline that legal evidence demands. • Driver-facing monitoring (DMS): distraction, eyes-off-road, drowsiness, phone use, and related unsafe-driving signals from the in-cabin camera. • On-device CV at the edge: getting detection models running efficiently on Kasava's Android dashcams (AOSP / CV-series) within real latency, power, and thermal budgets — not just in a notebook. • Environmental context: integrate HERE Technologies (and similar) to enrich events with weather, road, and location conditions — both continuously for risk context and precisely at the moment of impact — producing the verified record that feeds the signed claim packet. • A real evaluation and labelling loop: training, validation, and continuous improvement against messy real-world footage, with metrics that hold up under scrutiny. WHAT YOU BRING • 6+ years in software / AI engineering, with genuine production computer vision on video — object detection, tracking, scene understanding, and/or action recognition — not just image classification or LLM work. • On-device / edge ML in production: quantisation, ONNX, NPUs/DSPs, latency and power optimisation, and comfort getting models running on real hardware (Qualcomm SNPE / TensorRT or similar). • Sensor fusion instinct — combining video with IMU / accelerometer / GPS to corroborate events and cut false positives. • Rigorous evaluation: handling rare events and class imbalance, hard-negative mining, precision/recall trade-offs, and measurable robustness across night, glare, weather, occlusion, and mounting variation. • Strong Python and PyTorch and/or TensorFlow, plus the data-pipeline craft to build and maintain labelled datasets. • Fluent written and spoken English for a UK-headquartered team. BONUS POINTS • Driver-monitoring (DMS) or ADAS experience specifically. • Automotive, telematics, insurance, or fleet domain experience. • HERE / mapping / geospatial APIs and working with environmental/road-context data. • Deploying on Qualcomm dashcam SoCs and working close to the camera/ISP stack on Android / AOSP. • Active learning, auto-labelling, or edge-dataset tooling at scale. • Comfort handing clean, confidence-scored events to a downstream agentic claims pipeline — you'll partner closely with the AI backend team. WHY KASAVA You'll own the detection layer end to end — from the model to what runs on the camera to the verified incident record that becomes evidence. It's a rare problem space: computer vision, cryptography, and insurance, behind patent-pending technology, where what you ship has to hold up under real legal scrutiny. Room to grow a team as Kasava scales. Based in Kolkata, working with a UK-headquartered team.