POSITION DESCRIPTION

Data Scientist

Dashcam-based Pavement Defect Detection Programme

Function

In-house AI and Data Analytics

Number of positions

One

Reporting to

Head of Data Analytics

Engagement

Full time

1. OVERVIEW OF JOB REQUIREMENTS

1.1 Dashcam survey reports form the evidentiary basis for enforcement of operations and maintenance rectification. At present, third-party service providers both process survey data and report on the quality of their own outputs. The organisation is accordingly establishing an in-house defect detection model on AWS SageMaker, to provide an independent ground truth for measuring vendor precision and recall, and to retain direct control over model behaviour, tuning and interpretability.

1.2 The appointee will be the sole technical owner of this model and will be accountable for detection quality throughout the programme.

1.3 The mandate comprises four phases, as set out below.

Phase

Indicative duration

Scope

Data preparation

15 to 30 days

Establish video ingestion. Build the Ground Truth labelling workflow with machine-assisted pre-labelling. Define the annotation taxonomy and controls.

Model finetuning

3 to 4 months

Finetune a pre-trained detector on pavement defect data. Determine optimal configuration, validate across diverse geographies, and fix the evaluation protocol.

Model refinement

Continual

Deploy and orchestrate the retraining cycle. Validate vendor reports against model output, monitor drift, and provide interpretability for root-cause review.

Stabilisation

Approx. 1 month

Consolidate into a standalone report generation system. Transfer runbooks, model documentation and retraining schedules.

1.4 The system is to be designed for approximately 800 survey projects per week and approximately 137,000 defects per day requiring annotation at peak. Raw video is held at petabyte scale in cloud object storage,



and the model must generalise beyond the road conditions represented in the training data.

2. REQUIRED TECHNICAL SKILLS AND CAPABILITIES

2.1 AWS SageMaker: Demonstrated hands-on experience is required in the following services:

(a) Studio and Studio Notebooks, as a development environment;

(b) Ground Truth or Ground Truth Plus, for construction and administration of labelling workflows;

(c) Managed Training, Automatic Model Tuning and Experiments, for training execution, hyperparameter optimization and trial tracking;

(d) Managed Deployment, for configuration of inference endpoints;

(e) Pipelines, for orchestration of the retraining cycle; and

(f) Model Monitor, for detection of drift and degradation in deployed models.

Working familiarity is additionally required with Processing, Data Wrangler, Feature Store, Clarify, Debugger, Distributed Training Libraries, Managed Spot Training and bring-your-own container patterns.

2.2 Machine learning and computer vision:

(a) Code CI/CD experience production standard, with PyTorch or TensorFlow;

(b) Knowledge of object detection architectures, including transformer-based detectors such as DETR and DINO and convolutional baselines, with the ability to justify architectural selection on documented evidence;

(c) Experience in handling video and image processing at volume, including frame extraction, augmentation strategy and handling of GPS-tagged frames; and

(d) Expertise in evaluation methodology, including precision and recall trade-offs, intersection-over-union thresholds, mean average precision,



treatment of class imbalance, and selection of operating thresholds.

2.3 Cloud and engineering discipline.

(a) AWS foundations, comprising S3, identity and access management, storage lifecycle policies and cost control of GPU workloads; and

(b) Version control under Git, with dataset and experiment versioning sufficient to reproduce any published result on request.

3. OPTIONAL TECHNICAL SKILLS AND CAPABILITIES

3.1 The following skills are desirable to have:

(a) Condition assessment of infrastructure or civil assets, including pavement condition surveys and anomaly detection on physical assets;

(b) Experience in active learning or semi-supervised methods for the reduction of annotation effort;

(c) Integration of SageMaker with Kubernetes or Kubeflow, and configuration of multi-model endpoints;

(d) Knowledge in explainability methods for detection models, including saliency and attribution at frame level; and

(e) Fluency in data residency and government cloud compliance requirements applicable in India.

5. PAST EXPERIENCE

5.1 Technical Experience

(a) Minimum four years applying machine learning in production settings, of which at least two years on computer vision in object detection, segmentation or video understanding;

(b) End-to-end ownership of at least one model, from dataset design through deployment and post-deployment monitoring; research-only or infrastructure-only experience will not be accepted;

(c) Management or enablement of a human annotation workflow at scale, including quality control of labelled output;

(d) Delivery of a machine learning system in production on AWS SageMaker.

5.2 Educational Qualification

Bachelor's or Master's degree in Computer Science, Data Science, Electrical Engineering, Statistics or a cognate quantitative discipline. Equivalent experience may be considered in lieu.

📌 Data Scientist (Delhi)
🏢 Indypay Technologies
📍 Delhi