Recursio Chat
LLM interface + Tiny Recursive Model reasoning engine
System architecture
- 01
LLM interface
Conversational chat — Sudoku, mazes, ARC-AGI
- 02
Interface contract
A single contact point between chat and engine
- 03
Engine V1
Verifiable symbolic backtracking solver
- 04
Engine V3
Neural TRM-lite module, swappable with V1
- 05
Persistence
FastAPI + SQLAlchemy on PostgreSQL
Recursio Chat marries a conversational LLM interface to a reasoning engine built on Tiny Recursive Models (TRM). The V1 architecture rests on a symbolic backtracking solver, designed from the start so the neural TRM-lite module can cleanly replace it in V3 through a single interface contract.
What was built
- Single interface contract: the V1 symbolic solver and V3 neural module are swappable
- Symbolic backtracking solver for Sudoku, mazes, and ARC-AGI-style tasks
- FastAPI + Pydantic + SQLAlchemy backend on PostgreSQL
- pytest tests and Docker Compose orchestration from day one
- Philosophy: simple, compact structure — no premature abstractions
The real challenge
The design bet: start with a verifiable symbolic solver, but freeze the interface so cleanly that swapping in the neural (TRM) module touches a single component. Build for replacement, not for permanence.
Next project
Driver Fatigue Detection