Perfect Pair
An agentic returns assistant built for ALDO. It resolves customer returns, supports store decisions and investigates why products come back.
prize winner- Role
- Owned the full technical build on a team of five
- Context
- BSRM x ALDO Retail Gen AI Hackathon, October 2026
- Stack
- Claude on AWS Bedrock, Node.js, SQLite, EC2
- Outcome
- Selected as a finalist, then won a prize at the live final.
Three workflows
- Customer return
- Store decision
- Pattern investigation
Server
- Case state
- Tool execution
- Approval checks
- Saved replays
Model
- Claude on AWS Bedrock
In action
The problem
The hackathon theme was returns. A return costs a retailer twice: once to process, and again when the same product keeps coming back for the same reason.
What I built
- Three workflows in one interface: resolving a customer's return, supporting a store decision, and investigating return patterns across products.
- The server owns case state, tool execution and approval checks. The model proposes an action and the application decides whether it runs.
- Every session is saved and can be replayed, so anyone can see exactly what the agent did and why.
- Runs on Claude through AWS Bedrock. The hackathon copy was hosted on EC2 behind an IAM role.
- Works from synthetic operational scenarios plus aggregates of the data ALDO supplied. The raw files stay out of the repository.
Results
- Selected as a finalist after the first phase.
- Won a prize at the live final on October 15 and 16, 2026.
Limits
It is a demo, not a production service. Financial, inventory and publication actions are simulated.
Role switching is a demonstration control, not real authentication.
