Streamlining Image Review for a Financial Services Platform in the Auto Repossession Industry

Client Overview
The Challenge
The client was receiving thousands of vehicle images daily, each requiring manual review to:
Verify image completeness (e.g., front, rear, interior, etc.)
Confirm visibility of key vehicle identifiers
Assess damage
This process was time-consuming, error-prone, and not scalable, especially given the inconsistent image quality and the unpredictable nature of real-world submissions.

Key Constraints
Inconsistent Image Quality: Lighting, resolution, and angles varied widely, making standard analysis methods unreliable.
No Labeled Data: Lack of annotated datasets made it difficult to jumpstart a supervised learning pipeline.
Cross-Domain Complexity: No single technique, whether rule-based, GenAI, or deep learning, could solve the problem alone.
Unpredictable Edge Cases: The system needed to adapt to unforeseen image scenarios without constant reengineering.


Our Approach
We adopted a phased, experiment-driven development cycle:
Exploratory GenAI: Initial trials used generative AI to extract insights from vehicle imagery. While insightful, they lacked reliability and struggled with edge cases.
Prompt Engineering Refinement: Enhancing GenAI output through optimized prompts brought moderate improvements, but didn’t fully solve consistency issues.
Rule-Based & CV Integration: We layered in rule-based logic (for angle validation and metadata) and trained classical computer vision models for object recognition, improving precision but still limited in generalization.
The Hybrid Solution: Our final architecture combined the best of all approaches:
- Rule-based image angle validation
- Classical computer vision techniques
- Foundational mathematical logic
- Object detection & segmentation models
- LLM-based reasoning
This multi-layered pipeline worked in harmony to deliver real-time, zero-human-touch vehicle image analysis across thousands of images, with high accuracy and robustness.
The Outcome
100% Automated Image Processing
Real-Time Results with Zero Human Intervention
Scalable System Capable of Handling Thousands of Inputs Daily
Consistent Accuracy Across Varying Image Types and Edge Cases

What Made This Work
No Silver Bullet: A hybrid solution using diverse techniques was the only way to handle the problem end-to-end.
Iterate to Innovate: Each failed approach led to deeper understanding and more effective design iterations.
Start Simple, Grow Smart: Early GenAI exploration provided quick wins. Later stages improved inference costs, accuracy, and reliability.
Designed for the Unknown: Building flexibility into the system enabled it to gracefully handle new and unseen challenges without requiring a redesign.

At the heart of this success was our ability to fuse AI with real-world practicality, and that’s exactly what we do best. If you're facing complex data, image, or process challenges, we’ll help you turn them into scalable, intelligent solutions that just work.
Got a messy data challenge? We’ll help you clean it up with smart, scalable AI.CASE STUDIES