Computer Vision Engineer (2+ years)

The roleComputer Vision Engineer · Full-time · Indore (office-based) · 2–4 years' experienceYou will design, build, deploy and improve the vision models behind our products. This is not a research role: your work runs 24/7 on real cameras, in real conditions, for real customers.What you'll do1. Build models for production. Build vision models that run reliably in real operational environments, from manufacturing plants to offices, not just in the lab. Handle what real sites throw at you: changing light, motion blur, occlusion and awkward camera angles.2. Design and implement full pipelines. Own the whole lifecycle: data collection, labelling, model selection, implementation and optimisation. Then deploy to different hardware setups, from edge devices to on-premises servers.3. Work with real camera setups. Integrate and tune network (IP) cameras, industrial cameras and NVRs, and get reliable video from them into your pipelines.4. Extend our platform. Integrate new models, features and enhancements into our existing products without breaking what already runs.5. Scope new problems. Turn new customer problem statements into clear scopes, and build quick proofs of concept to test them.6. Build agentic workflows around detection models. Combine CV detections with LLMs and VLMs so the system can explain what it saw, decide what to do next, and trigger the right action. This is where we are heading; prior experience is a plus, not a requirement.What you bringMust-have1. 2–4 years of hands-on CV work in real-world environments, including deploying and managing CV solutions in production: rollouts, updates, monitoring and uptime across customer sites.2. Camera and stream handling: experience with either RTSP streams from IP cameras and NVRs, or industrial cameras (e.g. GigE or USB3 Vision, triggering, exposure and focus control). Plus camera calibration and setting up and managing regions of interest (ROI).3. Strong Python, deep learning and classical vision fundamentals:

• Production-quality Python: clean, modular, tested code; handling video at speed with threading or multiprocessing
• OpenCV for video input and output, frame processing and visualisation
• PyTorch or TensorFlow: custom datasets and data loaders, training loops, fine-tuning pretrained models, choosing and tuning loss functions
• Classical image processing and geometric vision: filtering, thresholding, morphology, contours, feature matching, homography and perspective correction, background subtraction
• Knowing when a classical method beats a deep learning model, and combining the two4. Object detection, segmentation and tracking with YOLO or similar models.5. Deployment readiness: adapting models for different deployment targets through quantisation (INT8/FP16) and other optimisation, plus MLOps basics (experiment tracking, model versioning).6. Measuring and improving models with mAP, precision, recall, FPS and latency, and explaining those numbers for your own projects.Good to have1. NVIDIA Jetson devices and the wider NVIDIA ecosystem (TensorRT, DeepStream, Triton, Metropolis, TAO Toolkit).2. Vision-language models (VLMs), e.g. to describe or explain what a camera sees.Know already, or learn quickly1. Backend skills: FastAPI, REST APIs, request handling, and database CRUD operations.2. CUDA basics and resource optimisation: profiling and tuning GPU and CPU usage (memory, batching, load balancing across streams) so systems get the most out of the hardware they run on.3. Managing remote deployments: accessing systems over SSH, pushing updates and troubleshooting at customer sites without being there in person.Why usA new direction, backed for the long term. Computer vision is a growing line of business at Vinfotech, and we are investing in it deliberately: more people, better hardware, and a widening set of problems to solve.Where we're heading. Our systems already run every day at customer sites. Building on them, we are moving from single-task detection models to vision systems that understand and act.How you'll grow:
• End-to-end ownership. You run systems from data collection to production, so you learn the full stack, not one slice of it.
• A path to specialise. Go deep in the area that fits you: edge deployment and the NVIDIA stack, industrial inspection, or VLMs and agents.
• Real deployments. Your work runs at customer sites, and you see its impact directly.
• Exposure to real-world problems. You work on live operational problems across different industries, and you solve them under real site conditions rather than with clean lab data.