The Problem with 'Stupid' Chatbots
Every founder has tried a "chatbot." They're usually just fancy versions of FAQ pages. They can't do anything. They just point users to documentation.
We wanted to build an Agent. Something with agency.
The Builder Stack
To avoid the "as a language model" trap, we focused on three technical pillars:
- Hyper-Specific RAG: Using Pinecone and OpenAI embeddings, we didn't just index docs; we indexed the internal Slack support history of the last 2 years.
- Native Tool Calling: We gave the agent
readandwriteaccess to Shopify's Admin API. It doesn't tell you where your package is; it checks the tracking, identifies delays, and issues a discount code if it's late. - Convex for State Management: Using Convex allowed us to keep the conversation state global and reactive. No complex Redux or separate DB sync needed.
The Build Log
The core logic was built in 4 days. The rest of the "14 days" was spent internal-testing and edge-case hardening (e.g., "Don't issue a refund for the same order twice").
The result: 82% of tickets are resolved without a human ever looking at them. For a builder, this isn't just a cost saver—it's a product moat.