AI Platform for automated road condition assessment using computer vision

Shepherd Services operates road asset management systems that rely on large volumes of road imagery. Traditionally, identifying defects such as potholes, cracking, surface wear, and line fading required manual review or inconsistent interpretation.
Whitefox was engaged to build an AI system that:

• Automatically detects road defects from captured imagery

• Classifies each defect according to a domain-specific taxonomy

• Measures defect magnitude and condition

• Assigns a consistent, deterministic severity rating (1–5 scale)

• Outputs structured, standardised data for integration into the RACAS® platform

We built a system that looks at road images, finds defects, evaluates how serious they are, and converts that into structured condition scores that engineers can use immediately.
The goal was not just detection, but repeatable, measurable, production-grade road condition intelligence at scale.
To deliver production-grade machine learning systems that convert unstructured data into structured, decision-ready intelligence, Whitefox built robust AI pipelines, validated model engines, deterministic scoring systems and cloud-native inference infrastructure, transforming raw data into trusted operational outputs.

Turning Raw Data into Production-Grade AI

Organisations often capture large quantities of high-value operational data (e.g., imagery, telemetry, logs, sensor streams) but struggle to translate it into actionable, consistent, real-world insight.
Common barriers to AI success include:

• Unstructured data that cannot be automated

• Generic detection models that lack domain alignment

• Lack of deterministic scoring or business logic

• Weak validation frameworks that don’t scale

• Infrastructure that fails under production load

When Shepherd Services engaged Whitefox for the RACAS® platform, these challenges were evident: vast imagery datasets with no automated intelligence layer and no scalable way to standardise severity or condition outputs across tens of thousands of records.

Whitefox was brought in to transform this into a reliable, repeatable, production AI pipeline.

Engineering the Production AI Pipeline

We apply a structured AI engineering discipline — combining machine learning, software infrastructure and operational rigour.

Step 1: Problem Definition & Domain Alignment

We worked closely with Shepherd’s product and technical teams to define:

• What “useful output” looks like

• What engineering standards to align to

• What performance targets are required

• What integration endpoints must be supported

This reframed the problem from “build a model” to “ship a production system.” We follow the intent-first mindset in building AI solutions, starting with what decisions users actually need to make, then engineering AI systems backwards from that intent.

Step 2: Data Transformation & Detection Pipeline

Raw unstructured inputs (e.g., imagery) were transformed into machine-readable formats using an automated AI pipeline:

• Image preprocessing

• Object detection architecture (YOLO format)

• Bounding box generation

• Confidence scoring and filtering

• Structured export for downstream systems

This pipeline eliminated manual review and enabled continuous processing of new data.

Step 3: Engineered Classification Ontology

Rather than adopting generic labels, Whitefox built a hierarchical, domain-aligned classification structure, including:

• Primary defect / feature classes

• Secondary groupings based on risk or category

• Explicit noise filtering classes

• Confidence thresholds for model outputs

This ensured that machine predictions were interpretable and actionable by domain experts.

Step 4: Deterministic Scoring & Rating Engine

Detection alone does not equal intelligence.

Whitefox developed a custom rating engine that:

• Converts model confidence into weighted scores

• Applies engineered severity tables

• Converts magnitude into 1–5 severity bands

• Produces deterministic outputs for operational use

This means the AI outcomes were repeatable and defensible — not subjective.

Step 5: Validation & Continuous Improvement

We implemented a rigorous validation framework that:

• Tracks performance by class (precision, recall)

• Expands training data intelligently

• Applies continuous retraining workflows

• Tunes confidence thresholds for production use

This framework scaled from tens of thousands to hundreds of thousands of labelled examples while maintaining performance integrity. This validation framework incorporates best practices from Deep Dive into DeepSeek for Software Engineers, including structured evaluation routines and metric tracking that support continuous learning and model interpretability.

Step 6: Cloud-Native AI Deployment

AI at scale requires more than models — it needs infrastructure.

Whitefox designed and delivered:

• Cloud inference clusters

• Data transformation services

• API integration layers

• Secure, hosted environments

• Monitoring and logs for production AI performance

This ensured the system operated reliably at scale without manual intervention.

Key Outcomes from the RACAS Engagement

While the context was road asset management, the underlying AI engineering strategy is domain-agnostic.

• Automated feature detection at scale

• Domain-aligned classification ontology

• Confidence-weighted severity scoring

• Deterministic rating engines

• Production validation framework

• API and infrastructure integration

This reflects Whitefox’s deep AI engineering domain expertise, applicable across use cases and industries.

Technical Highlights

Machine Learning & Computer Vision

• Custom object detection pipelines

• YOLO-derived architectures

• Confidence calibration and filtering

• Label hierarchy and ontology design

Model Governance

• F1 performance tracking

• Precision/recall balancing

• Class-level error analysis

• Continuous retraining workflows

AI Infrastructure

• Cloud inference clusters (AWS, GCP)

• Managed batch and streaming pipelines

• API integration layers

• Monitoring and logging at enterprise scale

Operational Integration

• Export-ready structured data

• Dashboard compatibility

• API hooks into workflows

• Compliance-ready outputs

Why Whitefox

AI succeeds in production only when models, data, infrastructure, and governance are engineered as a single system. This systems-first approach builds on our intent-first product philosophy, where AI is designed around real operational decisions, not abstract model performance.
Whitefox approaches AI as a full-stack engineering discipline. We design robust pipelines that operate reliably in real environments, validation frameworks that scale with growing datasets, deterministic scoring engines aligned with business logic, and cloud-native infrastructure built for continuous operation. Our integration layers ensure AI outputs flow directly into operational workflows, not isolated dashboards.
This systems-first approach is why our work moves beyond experimentation. We don’t deliver standalone models: we architect production AI platforms organisations can depend on.

Partner with Whitefox

Production AI requires more than experimentation. It demands structured platform engineering, robust system architecture, deterministic logic, and infrastructure that performs under real-world conditions.
If your organisation is working with complex, unstructured data and needs reliable, scalable intelligence - Whitefox can architect and deliver the full production AI system. From machine learning pipelines and scoring engines to validation frameworks and cloud deployment, we design solutions that integrate seamlessly into operational environments.