I didn't start with a plan. I started with a problem: every AI assistant I tried either lived in someone else's cloud, forgot everything between sessions, or couldn't be trusted with anything sensitive. So I decided to build one that didn't.
What followed was nine months of Python, dead ends, and hard lessons. The project has a name now — 7 — and a job description lifted straight from science fiction: a local-first, privacy-absolute executive assistant capable of anticipating needs, managing tasks, and improving itself over time. Jarvis without the cloud dependency. A Right Guide, not a search engine.
7 runs entirely on local hardware. No cloud routing. No subscription. No data leaving the machine. The architecture is a LangGraph orchestration pipeline — a supervisor that reads every request, routes it to the right specialist agent, and assembles the response.
A RAG layer called PANIC holds the knowledge base: embedded document chunks covering robotics, systems design, legal records, and operational doctrine. Memory write-back means 7 learns across sessions. A supervised self-modification loop means 7 can propose changes to its own codebase. The human approves. The system applies.
The architecture is built. The persona is constrained — an 8-billion parameter model on a ThinkPad can carry the routing and retrieval but not the full voice. That changes when the Beast arrives: Ryzen 9 9950X, 64GB DDR5, AMD GPU with 24GB VRAM. Two strings change in a config file. Everything else is already built and waiting.
You don't need a computer science degree. You need a clear problem, a precise specification, and the discipline to stay in the director's chair while the AI does the coding. The hardest part wasn't the code — it was knowing what to build and in what order. That's still true.