03USNov 2025 – Jun 2026
Modular AI Agent Platform
A production AI platform with a RAG chatbot and an architecture built to take on agents and business workflows without rework.
- steps to add a tool
- 3
The hard parts
The hard partarchitecture that survives growth
Service-oriented, with chatbot, retrieval, embedding, and agent services separated. Adding a tool is three steps: build it, test it standalone and against the LLM, register it with the right sub-registry. Nothing else in the codebase needs to know.
On multi-agent design
Passing every tool and document to one agent works in a demo and degrades in production. A routing agent reads intent and delegates to specialised agents with focused toolsets, over a shared knowledge layer. It costs an extra model call in latency and is harder to debug — so for a small assistant, one agent with a handful of tools is still the right answer. Knowing which situation you’re in is the actual skill.
Stack
- Python
- FastAPI
- PydanticAI
- vector search
- MCP