Challenge
A new Cocoon feature was needed to turn 40-90+ page pre-purchase home inspection reports, formatted differently by every inspection vendor, into consistent recommendations. Competing with quickly-built AI wrapper apps, it needed a pipeline accurate enough to generate trust with homeowners, yet sufficiently low-cost to protect margins.
Solution
Cocoon came to ATG for AI architectural guidance. ATG found the planned Textract and Comprehend approach couldn’t handle varied report formats and redirected the project to a multi-agent pipeline running Anthropic’s Claude on Amazon Bedrock AgentCore. The advisory engagement quickly became a working pipeline that extracts findings, assesses severity, and generates prioritized recommendations.
Duration: 6 weeks · Phases 4 [Design, Build, Optimize, Test] · Team: SA + CEM + ENG
Key Components
- Redesigned the architecture before building, replacing a costlier Textract + Comprehend plan with multi-agent Claude Sonnet on Bedrock AgentCore
- Processed report sections in parallel with Step Functions, so long reports finish in minutes
- Built in traceability and human review: every finding links to its source text and low-confidence findings pause for human-in-the-loop review based on adjustable thresholds
- Cut cost per report by 72% with prompt caching, from about $3.60 to about $1.00, with no loss in accuracy
AWS Services Used
- Amazon Bedrock
- AWS Step Functions
- Amazon Bedrock AgentCore
- AWS Lambda
- Amazon S3
Business Outcomes
- $0.02 projected cost per page in production
- 90–95% extraction accuracy on sample reports
- 72% lower cost with prompt caching
- <10 min to process a 70-page report
- ~$1.00 cost per report after optimization
- 98% faster than 8-hr target
Strategic Impact
Competing with quickly-built AI wrapper apps, Cocoon needed a pipeline accurate enough to generate trust with homeowners, yet sufficiently low-cost to protect margins. With ATG, the pipeline powers a new pre-purchase inspection feature, tightly tied to Cocoon’s core product, with traceability and human-in-the-loop capabilities needed to deliver the accuracy and reliability its stakeholders demand, at a $0.02 projected cost per page in production. Cocoon’s engineering team connects the pipeline to its platform, with ATG available for support.
