This Local AI Assistant Lives in a Suitcase

People usually choose a specific AI assistant because it helps them with the things that are important to them. When it’s not helping, it should disappear into the background until it is called upon again. But that isn’t everyone’s ideal assistant. Jim Kunz, for instance, is a maker who wants an AI assistant that is always present and always ready to interject its own opinions, whether they are helpful or not.

Kunz’s requirements were specific enough that nothing on the market could possibly make him happy. For this reason, he spent several weekends hacking together his ideal AI assistant that he named Sparky. It was built into a suitcase to make it portable — in the same way that an Osborne 1 is "portable." But while it may be bulky, that space is needed to run the AI algorithm offline and include all of the hardware that gives Sparky its personality.

Sparky is powered by an NVIDIA Jetson Orin NX Super 16GB single-board computer. The system hosts Google’s Gemma 4 E4B language model locally through llama.cpp, with response latencies as low as 200 milliseconds thanks to aggressive optimization of the model’s prefix cache. Unlike cloud-based assistants, Sparky operates entirely without internet access, APIs, or external servers.

Mounted inside the suitcase with the Jetson is an 11.6-inch HDMI display that renders Sparky’s animated face using a custom PixiJS interface running in a Chromium kiosk. The robot’s eyes track movement, the brows react dynamically, and its mouth synchronizes to speech generated by Piper TTS. Audio input is handled by a USB microphone and transcribed locally with SenseVoiceSmall speech recognition.

The Jetson and display are included in the Elecrow Jetson AI Starter Kit, as are more than 30 additional sensors that measure temperature, humidity, light levels, motion, distance, RFID presence, and orientation data. Rather than treating those readings as separate telemetry, Kunz injects them directly into the model prompt for additional context. With that information, Sparky might comment on the lighting in a room, notice that someone has moved closer, or mention the weather conditions around the case.

To get fast response times, Kunz had to do a lot of tuning. For instance, he found that moving constantly changing environmental information out of the system prompt and into the user message dramatically improved cache reuse, reducing cold-start delays from several seconds down to a few hundred milliseconds.

Power comes from a 50,000mAh battery pack hidden beneath the lid, allowing the robot to operate untethered for hours. Despite all of the work that went into the build, Sparky is already being treated as a stepping stone. Kunz says his next project will be based around NVIDIA’s AGX Thor platform with far larger memory capacity, enabling persistent memory, deeper vision processing, and more autonomous behavior. Stay tuned to Hackster News to see where this project goes next!