A private, fully self-hosted AI super-assistant. Your data, your hardware, your rules.
Project Jarvis builds a private, fully self-hosted AI super-assistant inspired by Tony Stark's J.A.R.V.I.S. The goal is a deeply personal AI that understands its user, integrates with the full fabric of their digital and physical environment, and acts as a trusted extension of their own capabilities.
The system is built entirely on open-source and self-hosted solutions, ensuring complete data sovereignty. No user data, context, or actions ever leave the local infrastructure unless explicitly initiated by the user through controlled channels.
Jarvis uses a coordinator + specialists pattern. Rather than one monolithic model trying to do everything, Jarvis is the single intelligent orchestrator that understands intent and delegates work to purpose-built specialist agents — each fast, predictable, and safe by design.
Every capability is manually added, documented, tested, and human-validated before activation. The system grows iteratively: one specialist at a time, fully validated before the next.
Jarvis is the coordinator — it reasons, decides, and orchestrates. A multimodal reasoning model that parses intent, manages memory, and delegates to specialists.
Specialists are smaller, faster models with tightly scoped capabilities. They execute tasks without touching the reasoning loop — fast, predictable, and safe.
No data ever leaves the homelab. Hardware kill switches on all sensors, encrypted audit logs, and a monitoring VM Jarvis itself cannot see.
Implemented one at a time — each validated in daily use before the next begins
Controls and queries the home environment via HomeAssistant. The MVP role — validates the core ask/act loop in a low-stakes, reversible domain.
Answers questions from the open internet and personal archives. Semantic NAS search, cross-source synthesis, and multi-document retrieval via Qdrant.
Homelab administration: monitoring, diagnostics, configuration, and scripting. Full safety stack — VM sandboxing, secondary model review, explicit authorisation.
Calendar management, weekly planning, grocery lists, personal finance. Heavily dependent on the episodic memory layer for rich, personal context.