The AI assistant operates entirely on-premises for privacy and latency reasons. It's designed for conversational, context-aware interaction without requiring wake words. Speech is continuously processed through a low-latency speech-to-intent pipeline optimized to respond in under 150 ms due primarily to ffmpeg's recent integration of whisper (in C++). This allows for seamless commands issued while moving between rooms, such as adjusting lights, temperature, and appliance states in a single multi-step instruction. I also started experimenting with Bark and now have near-real-time text-to-speech trained on Mike Tyson's voice. It's awesome to wake up to Iron Mike screaming "Get out of bed you pussy ass white boy!", but the primary voice is Don LaFontaine's--I trained the model on over 100 hours of youtube videos featuring his voice.

Additionally, I've been considering investing in some really high-end thermal imagers, but apparently that shit is extremely expensive and is a spook aphrodisiac (real-time high-resolution thermal imagers could be used for missile guidance and other antisocial behavior). Nevertheless, tracking within the house is an on-going issue and one I seek to resolve. I really need Don LaFontaine to exclaim "In a world where one man's job interferes with quenching his dog's thirst, ..." when my dog's out of water or something. Or, it'll just activate a switch that powers a normally-closed solenoid valve until her bowl is full. The nice thing is that it's fully extendable to whatever I can envision.

The assistant integrates directly with my (*extremely* custom) security system, CCTV server, HVAC, smart lighting, and a wide range of chinkshit IoT devices via secure APIs to isolated & sandboxed VLANs specifically for my chinkshit IoT devices. It supports chained commands ("dim living room lights, lower temperature to 72, arm perimeter turret with weapons") and maintains session context across multiple requests.

All devices communicate over encrypted channels with mutual TLS authentication to the extent possible with chinkshit IoT devices and their chinkshit proprietary firmware. The system's architecture is designed for high availability and failover, ensuring that automation functions remain operational even during maintenance or hardware failures. Primarily, this is accomplished with a blue-green deployment strategy. But, I also spin up on-demand GPUs for some stuff.

Taking q's for a little while.

(http://www.autoadmit.com/thread.php?thread_id=5762572&forum_id=2...id#49186379)

Also, for "sensitive" operations (unlocking doors, turning on sprinklers/garden irrigation, operating appliances like my oven/sous vide, and disarming the security system) it repeats the instruction and requests confirmation of its read-back. For example, if I say "Unlock the front door," it will respond with "Confirm instruction to unlock the front door" and unlock the front door after I affirmatively respond within two seconds. This confirmation isn't necessary for locking doors, arming the security system without weapons, turning off sprinklers/irrigation systems, etc. There's also built-in failovers upon power outages (normally-closed solenoid valves + UPS system that gives me at least one hour to connect to generator power before gracefully shutting down + backup starlink & 5G internet connection)

Of course, I've only really used it for asinine shit like providing recipe instructions while I'm cooking, setting alarm clocks from bed, turning on/off lights, etc. Huge time saver, even in spite of the herculean development efforts.

(http://www.autoadmit.com/thread.php?thread_id=5762572&forum_id=2...id#49186539)

can you remote in and do stuff from your phone?

(http://www.autoadmit.com/thread.php?thread_id=5762572&forum_id=2...id#49186544)

For really complex shit, it makes API calls to OpenAI (this is actually one of the tools in the "instruct" model)

(http://www.autoadmit.com/thread.php?thread_id=5762572&forum_id=2...id#49186634)