Location: Noida

Experience: 1–3 Years

Role Overview

This is a hands-on computer vision and applied ML role. Defects get described to you in plain, non-technical language by people who aren't vision engineers. Your job is to turn each one into a real vision problem: what camera view is needed, what evidence proves the defect, what technique to use, and when the honest answer is "we can't inspect this." You then take the ones worth building all the way through to a working, evaluated model running against a real accuracy target.

Datasets come unlabelled. Defect lists come as plain descriptions, not vision problems.

Building both is part of the job. You will also write the Python services that carry frames,

detections, and results through the pipeline, so this is more than model work.

Key Responsibilities

Problem Definition & Feasibility

- Translate defect descriptions given in plain, non-technical language into a defined vision
- problem: required camera view, evidence needed, and technique
- Identify and document cases where a defect cannot be reliably detected from the available imagery, with clear reasoning
- Define success criteria for each defect class before any model development begins

Data & Annotation Operations

- Own annotation guidelines and class definitions; administer the annotation platform
- Review annotation quality through inter-annotator agreement checks and resolve edge cases
- Maintain dataset versioning and enforce a held-out evaluation split with no leakage between training and test data

Model Development

- Develop detection, segmentation, oriented-box, keypoint, and OCR models as required; no single architecture covers every defect type




- Define per-class operating thresholds rather than a single global cutoff
- Conduct rigorous evaluation on held-out data and maintain a documented failure-mode analysis alongside accuracy metrics

Edge Deployment

- Export and quantise models (ONNX / TensorRT) for edge GPU inference
- Establish latency, memory, and throughput budgets based on real frame rates and image sizes
- Implement tiled inference for high-resolution frames that exceed single-pass model capacity

Pipeline Engineering

- Develop and maintain production Python services covering frame ingestion, object tracking across a pass, event publishing, and result formatting
- Work within an asynchronous, typed, and tested codebase using message queues and Docker
- Read, extend, and debug existing services in addition to building recent ones

Field Reality

- Work with real-world images affected by motion blur, occlusion, poor lighting, rain, and dust; design systems that degrade predictably rather than fail silently
- Undertake periodic site visits to assess real deployment conditions

Required Skills

- Production-quality Python: typing, tests, packaging, git, and code that others can pick up and maintain
- PyTorch,



and at least one detection or segmentation model you took from data to a working,evaluated result
- OpenCV and classical computer vision: geometry, calibration, colour, blur, thresholding,and the judgement to choose between classical and deep-learning approachesper-class precision/recall, a list of failure modes
- Ability to design an evaluation protocol before training anything: a held-out split,
- Comfortable reading and modifying code you didn't type yourself, including AI-assisted code
- Docker and basic Linux

Good to Have

- OCR / scene-text recognition (PaddleOCR, docTR, TrOCR, CRNN or similar)
- Model export and inference optimisation (ONNX, TensorRT, quantisation)
- Experience administering an annotation tool (CVAT or similar)
- Basic video and codec knowledge (FFmpeg, hardware encoding)
- Industrial/GigE Vision cameras, GenICam
- Self-supervised pretraining (DINOv2, MAE) or anomaly detection (PatchCore)
- Multi-object tracking across frames
- Message queues (MQTT/AMQP) or event-driven services
- Any exposure to industrial inspection or manufacturing QA

Education B.E. / B.Tech in Computer Science, Electronics & Communication, Electrical Engineering,

or a closely related discipline.

What We Value

- End-to-end ownership, from a rough defect description to a deployed, measured model
- Claims backed by evidence and measurement
- Full accountability for every line of code you submit
- A clean handover: documented assumptions, a repeatable evaluation, code someone else can pick up

Pay: From ₹600,000.00 per year

Benefits

- Flexible schedule
- Health insurance

Willingness to travel:
- 25% (Preferred)

Work Location: In person

📌 Computer Vision Engineer (Noida)
🏢 Cosmictech Builders Private
📍 Noida