Skip to content
Back to portfolio

Production full-stack AI system

CreatorChat

An embeddable AI chat & search widget that answers from a creator's own content — with an evaluation layer that catches hallucination before users do.

My role:Lead engineer — product + AI reliability

View on The Handbook (opens in a new tab)

Visual result

CreatorChat on The Handbook.

The Handbook Outfits page with the CreatorChat assistant answering a request for black party dresses and showing product cards.
The production assistant answering a shopping question with relevant product cards.

Product

What it does.

CreatorChat is an embeddable, customizable AI assistant for publisher websites. It gives readers grounded answers from the publisher's own editorial and product catalogue, including citations, product cards and streamed responses.

System design

Architecture without the confidential internals.

  1. 01Widget
  2. 02Cloudflare authentication and tenant scoping
  3. 03FastAPI / PydanticAI agent
  4. 04Bounded MCP retrieval
  5. 05Weaviate candidate search
  6. 06Canonical Supabase hydration
  7. 07Reference validation
  8. 08SSE response

Production reliability

Boundaries the system can enforce.

  • Tenant identity comes from authenticated context, not model input.
  • Retrieval remains scoped to the correct publisher and source.
  • Weaviate supplies candidates while Supabase remains canonical.
  • Unsupported references and links are rejected.
  • Retrieval calls and retries are bounded.
  • The agent clarifies or fails gracefully when evidence is insufficient.
  • Tests, evaluations, Logfire traces and Sentry support production operation.

Engineering workflow

How I use Claude Code.

I use Claude Code for:

  • Codebase and architecture mapping
  • Bounded implementation planning
  • Implementation and targeted tests
  • Logfire trace investigation
  • Prompt-behaviour evaluation
  • Security, senior-engineering and skeptical review
  • CI diagnosis and evidence preparation

Human control

Architecture, database mutations, prompt promotion, deployment and rollback remain human-controlled.

Disclosure

What I can share.

CreatorChat and its integration with The Handbook are proprietary client work. I cannot publish the production repository or confidential implementation details. This case study describes only the architecture, responsibilities and engineering decisions I am permitted to share.