The 2013 Recon Instruments Jet "smart" sports glasses, which originally ran on Android, have been reflashed to Mainline Linux 6.18: the system boots in less than four seconds, and a voice AI assistant has been deployed on the device. The wake-word detector and lightweight models remain on the glasses themselves, while heavy inference—vision, speech recognition, language model, and text-to-speech—is handled by a home server. The author performed reverse engineering of the outdated hardware and driver porting alongside Claude Code. In the final build, the glasses describe scenes, hold chats, scan objects with bounding boxes, display OpenStreetMap maps, and even run Doom.

What happened

The author ported the Recon Instruments Jet glasses from the Android platform to Mainline Linux 6.18, achieving a boot time of less than four seconds. Drivers for the Kopin microdisplay with a 428x240 resolution, cameras, microphones, Wi-Fi, Bluetooth, IMU, and GPS are now functional. The computing architecture separates roles: only the microWakeWord wake-word model, which consumes less than 2% of the processor, remains on the glasses, while heavy inference—scene description, ASR, LLM, and TTS—is handled by a home server equipped with an AMD Ryzen AI Max 395+ "Strix Halo" and 128 GB of memory, where llama-swap routes between models. To make the assistant seem responsive, answer streaming, model pre-warming, and pre-prepared voice responses are applied. In working modes, the glasses describe the scene around the user, hold chats, and scan objects with bounding boxes in a "world model"; the device opens OpenStreetMap maps, provides shell access, and runs Doom via Chocolate Doom for entertainment.

Context

The Recon Jet is a 2013-generation sports "smart" glasses device featuring a TI OMAP4430 chip, 1 GiB of RAM, and 8 GB of eMMC. Devices of this age typically remain on abandoned Android firmware: porting to a current kernel requires reverse engineering of the bootloader and peripherals, which demands rare and expensive expertise in embedded systems. The distinction of this project is that a significant portion of this work was delegated to the LLM agent Claude Code, changing the premise itself: a path previously available only to narrow specialists was accomplished by a single enthusiast. The "light on device, heavy nearby" scheme belongs to the standard edge AI toolkit: it keeps voice and video data with the user and operates without cloud services, which is fundamental for private wearable assistants.

Why this matters for the industry

Two cost reductions are important for the industry in this case. First: the LLM agent actually took on low-level reverse engineering and driver porting for outdated hardware—a task that previously required rare specialists; this reduces the cost of reviving legacy devices and supporting embedded platforms. Second: the combination of a lightweight wake-word detector on a wearable device and multimodal inference on a consumer-grade home server demonstrated a working architecture for a cloud-free voice AI companion, suitable as a reference for private wearable products. For teams building edge assistants, the set of UX techniques used by the author to mask remote model latency holds separate value—such techniques can be transferred to commercial prototypes.

Why this matters for users

For the reader, this is a clear DIY example: "dead" sports hardware from thirteen years ago is transformed into a Linux platform with voice AI, a camera, and a "world model," and almost every element of the scheme can be built in a personal prototype—a local wake-word detector on weak hardware, heavy inference on a personal server, and an LLM agent as a helper for low-level porting. The honest experience of delegation is also useful: the author explicitly states that Claude Code produced "spaghetti code" and made diagnostic errors, so the agent's results must be manually verified. There is also a hard practical limitation: the glasses' battery lasts about an hour at a current of around 300 mA, so autonomy remains a weak point for such a device in its current form.

What is still unknown / limitations

Final conclusions cannot be drawn: only the first part of the series has been published in a personal blog, and the code and drivers have not been released, so the claimed port to Mainline Linux 6.18 remains unverified. There is also no independent verification—the discussion on Hacker News consists of one point and one comment. The report lacks measurable benchmarks: there is no data on end-to-end assistant latency, speech recognition accuracy, scene description quality, or "world model" bounding boxes, and the provided figures are given on the author's word without external measurements. It is unknown whether a second part of the series will be released or if the code will be open-sourced, and this determines whether the case will become a repeatable recipe or remain a one-off experiment.

Sources

Author

Look at AI, editorial team