The scenario
CartPilot is a Forward Deployed Engineer exercise. You inherit an e-commerce monolith you didn’t write, map how it works, and prove value with a standalone shopping assistant built on Pydantic AI and a fast open-weight model. Once the proof-of-concept passes its guardrail and quality checks, you integrate it into the main FastAPI application — catalogue search, cart actions and order lookups as typed tools — and deploy the unified service to Cloud Run with tracing and automated evaluations in CI.
System architecture
01 Storefront Chat widget
02 FastAPI monolith Existing routes + agent route
03 Pydantic AI agent Typed tools + guardrails
04 Catalogue & orders Postgres via the app’s services
05 Cloud Run Traced with Logfire
High-level flow of CartPilot, from input to output.
What you will learn
- Running a codebase discovery and writing an integration plan
- Typed tools for catalogue, cart and order actions
- Input/output guardrails for a customer-facing agent
- Evaluations that run on every pull request
- Observability with traces across the agent and the app
What you will build
- Discovery report and integration design for the monolith
- Standalone assistant with search, cart and order tools
- Integrated FastAPI service with a chat widget in the storefront
- CI pipeline with evals and deploy to Cloud Run
Build plan
- Discovery: map the monolith and pick integration points
- Prototype the assistant outside the codebase
- Add guardrails and an eval set
- Integrate into the monolith behind a feature flag
- Deploy with CI evals and tracing