AI Mesh (Delta Topology)

Overview

The P31 AI mesh uses a delta topology — three primary models plus a local fallback, routed by a semantic classifier. No single model is a dependency. The system degrades gracefully: if Claude is unavailable, queries route to Gemini; if all cloud models fail, Ollama handles everything locally.

Model Roles

Model	Role	Provider	Use Case
Claude (Sonnet)	Reasoning	Anthropic	Architecture, integration, complex analysis
DeepSeek Coder V2	Code	DeepSeek	Firmware, tests, code generation
Gemini 2.5 Pro	Vision	Google	Documentation, schematics, long context
Qwen 2.5 Coder 7B	Local	Ollama	Offline summarization, privacy-sensitive

Routing

The semantic router (backend/router.py) classifies incoming queries into domains using embedding-based similarity matching. Each domain maps to an optimal model.

Domains: FIRMWARE, FRONTEND, BACKEND, COGNITIVE, DOCS

When the semantic router is unavailable (first run, no embeddings), a keyword-based fallback provides basic routing.

LiteLLM Proxy

All model calls route through LiteLLM at :4000, which provides:

Unified API across providers
Automatic failover (configured fallback chains)
Request retry (3 attempts, 30s timeout)
Cost tracking

Configuration: .devcontainer/litellm-config.yaml

Continue.dev Integration

The .continue/config.yaml configures VS Code’s Continue extension to use the mesh:

File-type routing: .c/.h files → DeepSeek, .tsx/.jsx → Claude
Image/schematic queries → Gemini
Summarization → Local (Ollama) for privacy
Embeddings → transformers.js (runs in-browser)

Fallback Chain

reasoning → multimodal → local
code      → reasoning  → local
multimodal → reasoning → local

Every query has a path to completion, even fully offline.