Atlas + Jarvis
A voice-first personal AI operations system, self-hosted.
Atlas + Jarvis is a personal AI operations system that runs on a home server with on-device automation: a voice-first assistant that handles scheduling, task orchestration, and daily briefings. It pairs a low-latency conversational interface with a tiered model architecture, so quick questions answer instantly while heavier work runs in the background and reports back. Built and run daily as a working tool, it lets a surgeon keep every part of life moving forward despite an unpredictable schedule.
The problem
A demanding surgical schedule makes it easy to fall behind on everything outside the operating room — tasks, study habits, follow-ups, and the dozens of small threads that quietly slip. Existing tools were scattered and required constant manual upkeep; nothing connected capture, planning, and follow-through into a single loop. The goal was a system that keeps every pillar of life advancing without depending on willpower or streaks, and that is fast enough to actually use in spare moments.
What I built
A self-hosted personal AI operations system with two main surfaces: a voice-first conversational assistant and an automated life-planning engine. The assistant offers a clean, voice-only interface with streaming responses and natural speech replies, reachable privately from a phone. Behind it sits a planning layer that captures voice and text input, organizes tasks and multi-step threads, and produces daily briefings. A safety layer gates any outbound action, so the system can act autonomously on the owner's own data while never sending anything to another person without an explicit, separate confirmation.
How it works
The system runs entirely on a home server with on-device automation, exposed only over a private encrypted network. A tiered architecture routes each request by weight: a fast model handles quick voice turns, while heavier multi-step jobs are dispatched to a more capable model in the background and the results delivered back so nothing is lost. Speech is streamed and chunked sentence-by-sentence for near-instant playback, and the assistant keeps per-conversation memory so turns build on each other. Background rebalancing loops periodically scan commitments, classify them by life domain, and resurface what is being neglected rather than only what is loudest — surfacing slippage before it compounds.
Where it stands
Live and in daily personal use. Latency was tuned from tens of seconds per turn down to roughly two to four seconds through pre-warming, keep-warm, and streaming. The autonomous-action safety gate, audit logging with one-command undo, and the neglect-aware rebalancing loops are all built and verified. It is an actively maintained, working system rather than a prototype — relied on to capture, plan, and follow through across clinical, research, and personal commitments.