Computer Vision Engineer - Device Grading

Computer Vision Engineer - Device Grading

India (Remote)

Department:

Engineering - AI / ML

Reports to:

VP Engineering

Experience:

4-5 years

.

We are hiring a Computer Vision Engineer to own automated device grading - detecting and scoring cosmetic damage (scratches, dents, cracks, screen defects) from photos and video captured by customers on their own phones, in uncontrolled lighting, on highly reflective surfaces.

This is a founding computer vision role. You will be the first dedicated CV engineer at owning the problem end to end: dataset and annotation strategy, detector training and tuning, production inference, and the closed-loop feedback that improves the model with every device it sees. There is no existing ML platform to inherit.

Grading is a hard detection problem, not a demo. A hairline crack is a handful of low-contrast pixels; a ceiling-light reflection on glossy glass looks exactly like a scratch. We need someone who has fought these problems in production, not only on clean benchmark datasets.

WHAT YOU WILL DO

Detection Modelling

• Train, tune, and deploy YOLO-family detectors as the backbone of our grading pipeline - owning the full loop from label assignment and augmentation strategy through loss tuning, hyperparameter search, and export
• Solve the hard detection cases: small-object and low-contrast defects such as hairline cracks and light scuffs, and distinguishing genuine surface damage from glare and reflections
• Benchmark YOLO variants against alternatives (DETR-family, segmentation, anomaly detection) and justify the choice on measured evidence rather than familiarity
• Extend detection into severity - translating detected defects into the grade bands the business actually trades on

Data Annotation

• Own dataset strategy: collection, curation, class balance, and targeted capture for defect types we see too rarely
• Define the annotation rubric and measure inter-annotator agreement - grading labels are subjective, and model quality is capped by label consistency
• Build augmentation that reflects reality: varied lighting, glare, reflections, device colours and finishes, and the full range of customer capture quality
• Version datasets and keep a trustworthy held-out test set that reflects production conditions, not lab conditions

Production Inference

• Take models from experiment to production - export and optimisation (ONNX, TensorRT), serving, versioning, latency and throughput targets
• Reason about where inference runs: lightweight on-device models for capture guidance versus full server-side grading, and the trade-offs each way
• Instrument models in production: confidence distributions, per-device-model performance, and drift detection as new handsets enter the catalogue
• Partner with backend engineers to integrate inference into platform flows, including the escalation path when the model is not confident enough to decide

Closed-Loop Feedback

• Build and maintain a closed-loop system: capture production predictions alongside the corrections made when devices are physically inspected, and feed them back into training
• Prioritise which corrections are worth labelling - disagreements and low-confidence cases first - so annotation spend buys the most model improvement
• Run retrain, evaluate, and redeploy as a routine cycle with regression testing, not an occasional project
• Set the technical direction for computer vision at and write the design docs and evaluation reports that let non-specialists act on the results

WHAT WE ARE LOOKING FOR

Must Have

• 4-5 years of professional computer vision experience, with detection models running in production
• Extensive, hands-on YOLO experience - you have trained, tuned, and deployed YOLO detectors on your own data, not just run a pretrained model. You can discuss version trade-offs, anchor-based vs anchor-free assignment, NMS and confidence-threshold behaviour, and the pitfalls of exporting to a production runtime
• Demonstrated small-object and low-contrast detection work - the cases where a defect is a few dozen pixels and barely separable from the background
• Strong augmentation and hard-negative strategy for uncontrolled capture conditions: lighting variation, glare, reflections, motion blur
• Strong Python and PyTorch; comfortable owning the full pipeline from raw images to a deployed endpoint
• Proven production deployment - export and optimisation, serving, latency, versioning, and monitoring in a live system
• Has built and maintained a closed-loop feedback system where production corrections flow back into retraining
• Evaluation discipline: chooses metrics that reflect business cost, builds honest test sets, tunes thresholds for asymmetric error cost, and can explain exactly where a model fails
• Self-directed and comfortable as the first and only CV specialist - you will set the approach, not inherit one
• Able to work with backend engineers to get models into production, not hand them over the wall

Nice to Have

• Device grading, visual quality inspection, or industrial defect detection experience
• Anomaly and few-shot approaches (e.g. PatchCore) for defect classes with very few examples
• Instance or semantic segmentation to measure defect extent, not just presence
• Video and multi-frame assessment - frame quality scoring, aggregation across angles, guided capture UX
• Edge or on-device inference and mobile capture pipelines
• Serving and optimisation stacks: ONNX, TensorRT, Triton, or SageMaker
• Vision-language models applied to inspection or grading tasks
• MLOps tooling for experiment tracking and data versioning (MLflow, Weights Biases, DVC)
• Active learning or semi-supervised labelling to reduce annotation cost
• Recommerce, e-commerce, or marketplace domain experience

Tracker:

Vendor Name

Candidate Name

M.No

Candidate Email

Skill

Experience

Notice Period

Location

Rate Card