Fuel Logistics / Document Processing · AI/ML Engineering · AgentCore Quick Start · BUILDSTR
Challenge
Keep the pipeline moving as formats multiply
DashFuel runs a fuel logistics platform to help their customers maximize fuel investments. Drivers photograph bills of lading at more than 10,000 fuel terminals, hundreds a day, and every terminal uses its own format. DashFuel already read known formats well: Amazon Textract plus Claude on Amazon Bedrock had raised extraction accuracy from about 60% to 95–99%. New formats were the gap. A bill of lading from an unmapped terminal couldn’t flow through the pipeline until someone built a processor for it.
- Each new format stopped the pipeline for about 90 minutes while an engineer gathered samples, hand-built a processor, and validated it
- The work fell on key engoineering resources, which pulled them off core product development
- Onboarding effort grew in step with every new customer and its terminal coverage, across a field of roughly 15,000 formats
- Driver photos vary in quality, layout, and handwriting, which makes extraction harder
- The team was too busy delivering core features to the platform to address the gap
Solution
An agent that drafts the processor
In a four-week AgentCore Quick Start engagement, BUILDSTR delivered a human-in-the-loop agentic solution with Strands Agents on Amazon Bedrock AgentCore, deployed with Terraform into DashFuel’s own AWS account alongside its existing AWS Step Functions document pipeline. The agent works like a human engineer: it reads the repo, writes the code, runs the tests, and asks for review.
- Detect and classify: spots bills of lading from unmapped terminals and decides whether each is a known format or a new one
- Draft: writes a processor for each new format from DashFuel’s existing templates. Each document is routed to Amazon Textract for structured fields or to Claude on Bedrock for less structured layouts.
- Validate: runs the draft against DashFuel’s existing test harness before a human sees it
- Review loop: opens a pull request in DashFuel’s GitHub repo. An engineer reviews it, the agent refines it, and nothing ships without approval.
- Visibility: AgentCore-native observability traces every agent decision end to end. Operators use a web UI behind Amazon CloudFront with Amazon Cognito sign-in.
- Reuse: delivered as Terraform, with the architecture documented as a template for future agents
Result
A 90-minute build became a 5-minute review
By handoff, the processors the agent generated were passing DashFuel’s test harness, and the engineer’s role had moved from author to reviewer. DashFuel came away with a working agent in its own account, the Terraform to redeploy it, and a readout with production recommendations.
- 94% less engineer time per new format: about 90 minutes down to 5 minutes
- DashFuel can take on new formats across a field of roughly 15,000 without adding engineers
- Engineering hours go back to core product work that had been blocked
- A production-ready pattern and a repeatable agent approach on AWS
- Paved the way for further work on Tenant onboarding and invoicing agents
