Notes from the team.
Technical deep dives, open-source releases, and developer guides for on-device AI.
We put a voice agent on Genie's delivery robot. Then it rode away.
We shipped an on-device voice agent onto Genie Mobility's autonomous delivery e-bikes in ten days. The engineering worked on the first try. The UX lesson split the stack in a way we hadn't planned, and pushed us into the harder problem underneath: what the contract between a language model and an autonomy stack has to look like before you let one supervise the other.
Read→Matic proved people want to talk to their robots. The next frontier is where that voice runs.
Matic Cues brings voice and gesture control to the robot vacuum, with the wake word and the entire vision stack running on-device and command understanding going to Google's Gemini API in the cloud. Here's what that architecture says about the real state of on-device voice AI on a Jetson.
Read→Barge-in and interruption handling for on-device voice agents
Barge-in is the hardest part of a voice agent to get right: hearing yourself over your own TTS, deciding an interruption is real, and stopping in a couple of audio frames. Why the DIY Whisper + llama.cpp + Piper stack breaks it, and how an on-device C++ loop fixes it.
Read→Time-to-first-token: why voice agents live or die on latency
A latency budget for real-time voice agents: the 200ms human turn-taking rule, where cloud pipelines spend 600ms–1.7s, and how on-device buys it back.
Read→How to build an offline voice assistant on NVIDIA Jetson Orin
A builder's guide to running a fully offline voice agent on Jetson Orin — local STT, an SLM, and TTS, with real latency numbers and where the DIY Whisper + llama.cpp + Piper stack breaks down.
Read→Why we ripped cloud voice out of our robots: from Deepgram + OpenAI + ElevenLabs to on-device
We started building robot voice the easy way — Deepgram, OpenAI, and ElevenLabs glued together in Python. It demoed great and fell apart in the real world. Here's everything that went wrong, and why we're rebuilding the whole stack on-device.
Read→turboquant.cpp: near-optimal vector quantization in 400 lines of C++, no training required
We open-sourced turboquant.cpp, a C++23 implementation of TurboQuant: compress embeddings to 1-4 bits per coordinate with provable distortion bounds — no training, no codebooks.
Read→