Prism
Personal AI Router
Prism is the command layer I use to turn plain-English messages into useful work. It routes Telegram input, scheduled jobs, memory, connectors, and model execution into the right lane, then keeps enough context for the next decision to be easier. The point is not chat. The point is a personal operating system with proof.
Connectors
AI Router
Codex
Code + review laneRepo-grounded planning, implementation, verification, and review work where codebase evidence matters most.
Claude
Conversation + synthesis laneHigh-level planning, writing, summarization, taste checks, and broader reasoning loops where it is the better fit.
Local Router
Intent layerLightweight classification and command routing before heavier model calls are needed.
How It Works
Inbound
Messages arrive via Telegram (natural language or slash commands), scheduled cron tasks (daily briefings, market digest, weekly themes), or webhook triggers from external services.
Free-Form Routing
A local Qwen 2.5 7B classifier reads plain-English messages and decides: invisible slash command, sub-agent dispatch, or agent-loop question. Shadow mode + circuit breaker kept it honest for two weeks before it went live.
Cross-Model Execution
Codex and Claude are routed by job instead of treating one model family as the whole system. Code-heavy work, review, planning, synthesis, and conversational tasks each get the lane that fits.
Applied Intelligence Layer
Weekly themes digest ships every Sunday 9:30 AM EDT. Clusters the week’s activity across ingestion + writing + dispatch into a one-message synthesis. The layer that makes the portfolio feel like a single instrument.
Python, FastAPI, Codex CLI, Claude, local Qwen 2.5 7B (Ollama), Supabase, MCP servers, Cloudflare Tunnel, scheduled local runtime