Alikhan Makhamadaliev

Self-hosted agentic-AI daemon. It lives on my metal, remembers, and acts. It answers only to me.

Runs 24/7 on a Mac Studio under macOS launchd with a watchdog that self-heals the service. The asyncio agent kernel keeps a process table, runs up to 4 concurrent LLM agents, and schedules them on a 3-level multilevel feedback queue with per-level wall-clock allotments, demotion accounting, and a periodic priority boost so nothing starves.

Tool permissions are deny-by-default. Autonomous kernel agents get a read-only tool set behind a central capability manifest, read confinement, and SSRF egress denies, which closes the prompt-injection lethal trifecta. A dedicated tool-lock regression suite guards it.

runtime

Python 3.14, asyncio, FastAPI, SQLite

scheduler

3-level MLFQ, up to 4 concurrent agents

plugins

54 auto-discovered, behind an LLM intent router with a local Ollama shadow-routing tier and chat fallback

transports

iMessage, reading the macOS chat.db directly, plus a FastAPI web app over a Tailscale tailnet with bearer-token auth and a CIDR allowlist

backend

headless Claude CLI subprocesses on a flat subscription, no per-token API cost

tests

1,382-test offline unit suite, plus a tool-free LLM grounding judge that checks agent output against its sources

durability

crash-consistent atomic state writes, SQLite online-backup snapshots every 6 hours