Amazon Bedrock listing review agent helps Cars & Bids maintain quality at scale

Cars & Bids is an online auction marketplace dedicated to modern car enthusiasts, founded by automotive reviewer Doug DeMuro. The platform offers a curated environment where sellers can list "cool" vehicles, from high-performance sports cars to quirky trucks, for free.

With an 80%+ sell-through rate and a high quality standard on listings, Cars & Bids sought to improve the consistency and efficiency of their media management and listing review process without sacrificing the reputation for excellence they have with their customers. 

Challenge: Improving consistency and efficiency of time-intensive review workflows

Due to the high value of transactions, automotive listing review carries significant complexity and risk. Ownership records must be reviewed, submitted media must be checked for required views, quality issues, and documentation coverage, and key vehicle details must be validated to support listing quality.  

The team wanted to streamline its internal listing-review process and make submission, photo, and video review more efficient without changing its existing quality bar.

Efficiency mattered, but so did control. Human review and detailed observability at key stages was non-negotiable. 

Solution: AI powered media processing and listing review to enable scale

Tech 42 worked with Cars & Bids to scope and build a listing-review agent that would help Cars & Bids specialists evaluate seller-submitted information more efficiently. The agent reviews seller messaging history, vehicle-specific details, known issues, model-specific quirks, ownership/title gaps, and media review needs. Then it analyzes the data for flags, key questions, and next steps to guide the person working with the seller. 

Tech 42 built three interconnected systems to handle the full scope of the project.

Outcome: Benefits to both internal teams and customers

As of writing, the Cars & Bids team has delivered 36K+ completed auctions for a total value of $800M+ in vehicles sold. With a platform gaining traction with 1.1M+ members and growing, they need to be ready for maintaining quality at scale. 

Through the new AI-powered tooling, the Cars & Bids team is seeing benefits across three dimensions as they keep pace with growing volume: 

Benefits for Cars & Bids listing specialists

The complete system helps specialists review submission data, photo labels, video labels, notes, and ownership/title evidence in a more consistent workflow. It surfaces missing information, media issues, and suggested seller follow-up so the team can maintain its existing quality bar with less manual triaging.

Seller and buyer impact

A strong listing review process directly supports the platform's 80%+ sell-through rate and addresses the top customer support complaint: taking all the right photos is hard. Sellers can receive clearer follow-up requests sooner, and buyers continue to see listings that meet Cars & Bids' established documentation, media standards, and high quality bar. 

Benefits for the platform

The pipeline improves internal review scalability without requiring proportional increases in manual effort. Langfuse-powered observability gives the Cars & Bids team trace-level visibility into agent behavior and feedback, while Terraform-managed infrastructure supports reproducible deployments and environment teardown.

Technical details: Building agentic workflows on AWS

Tech 42 focused on building three interconnected internal systems to support listing review:

1. AI listing review agent

An Amazon Bedrock-powered LangGraph agent assists Cars & Bids listing specialists by reviewing structured seller submission data and surfacing vehicle-specific follow-up needs.

 It uses curated vehicle and model-specific CSV knowledge files to identify relevant follow-up questions, known issues, title/ownership concerns, and details that enthusiast buyers expect to see in a complete listing.

The agent returns structured JSON for downstream integration, including summary points, ownership verification signals, and suggested seller follow-up.

The agent is equipped with tools, Postgres-backed checkpointing/memory, and Langfuse trace/feedback observability.

2. Automated photo classification agent

A Lambda-based photo labeling system processes uploaded photos and produces structured label JSON for specialist review. The delivered system includes a Bedrock VLM path and a classifier-first ONNX path with LLM fallback for documents or low-confidence cases. It records labels, quality flags, scores, and classifier diagnostics to help the team maintain existing photo standards more efficiently.

3. Automated video processing agent

A Lambda-based video processing workflow handles seller-uploaded videos. Using TwelveLabs Pegasus through Amazon Bedrock, the system generates tags, summaries, and structured JSON artifacts for downstream review.

AWS architecture overview

The solution is built on AWS infrastructure, leveraging managed services and AI/ML capabilities for internal review, media processing, storage, deployment, and observability.

Technology stack

• Infrastructure as Code: Terraform
• Language: Python
• Agent Framework: LangGraph
• API Framework: FastAPI
• Compute: Amazon ECS
• AI/ML Platform: Amazon Bedrock for agent and multimodal model workflows; curated vehicle/model-specific CSV knowledge files
• Photo Classification: Bedrock VLM path, classifier-first ONNX path, and LLM fallback for document/low-confidence cases
•  Video Analysis: TwelveLabs Pegasus via Amazon Bedrock
•  Monitoring: Langfuse and CloudWatch
• Storage: Amazon S3, Amazon RDS PostgreSQL
• Serverless Processing: AWS Lambda

Internal review process