On-device inference + cloud LLM · OTA
AI Products
From a desktop robot to an AI toy core: devices built around a 0.5–16 TOPS NPU that handles everything up to the model — wake word, far-field audio, speech recognition, vision — with the language model in the cloud behind it.
The part that decides the project
Latency, and where you spend it.
A voice device is judged on how long it sits silent after you stop talking — under roughly 700 ms it feels like conversation, past 1.5 s it feels broken. With the model in the cloud, most of that budget is already spent before the network is touched: wake-word detection, endpointing, noise suppression and speech recognition all run on the device, and every millisecond they waste comes out of the model’s allowance.
OTA
OTA is in every connected design
Firmware ships with a dual-partition layout and automatic fallback, so an interrupted update boots the previous image instead of bricking the unit. Updates go out in staged batches from the device cloud, with version reporting per unit. It is designed in at the first prototype, not added before launch.
What you need to decide before we can quote
- How much runs on-device — wake word, full speech recognition, or a small local model
- Response budget: how long the device may take before the model is called
- Which cloud model sits behind it, and in which region
- What still works with no network — and what you promise the user
What this direction covers
- AI desktop robots
- In-car AI assistants
- AI companion devices
- AI toy cores
- Connected collectibles and board-game toys
- Bedside sleep and care companions
- AI camera modules
- AI drawing printers
Silicon it runs on
Linux SoC with NPU (0.5–16 TOPS) · WiFi SoC · MCU
CapabilitiesEngineering modules
Already proven in shipped projects
The building blocks a new project starts from, rather than from scratch.
Voice front end
- Offline wake word, user-defined
- Mic-array direction of arrival, with false-turn suppression mid-conversation
- Echo cancellation while the speaker plays
- Low-latency streaming, endpointing tuned to the latency budget
Motion
- Head pitch motor (nod, look up) and body yaw motor
- Continuous 360° turntables
- Micro servos inside plush — ears, limbs
- Motion quieted while the mics are listening
Sensing
- Multiple NFC readers in one body — accessory ID and tap-to-pair
- Capacitive touch zones classified by duration
- 6-axis IMU gestures: shake, tilt, pick-up
- mmWave radar presence and breathing; PPG heart rate and SpO₂
- Optional camera for face detection
Connectivity and cloud
- Wi-Fi / BLE with app provisioning; 4G Cat.1 with GNSS, combined or separate
- Vehicle link over Bluetooth, CAN or USB
- Works with your protocol and cloud, or ours
- Device behaviour orchestrated from the cloud
Expression
- Round LCD and dual-eye screens, LED dot matrices, RGB rings
- 100+ expression and sound assets
- Idle-behaviour scheduler: moods, naps, wake-ups
Voice dialogue
Voice dialogue, working today: 60 language variants
The dialogue chain is integrated end to end on our voice boards. The international deployment covers speech recognition and synthesis for the languages below; the China deployment adds a wide choice of Mandarin voices, plus Cantonese, Sichuanese, Henan and Hunan dialects. Each project picks the deployment region and voices for the markets it ships to.
Europe · 29
English (UK), French, German, Spanish (Spain), Italian, Portuguese (Portugal), Dutch, Polish, Russian, Ukrainian, Czech, Slovak, Hungarian, Romanian, Bulgarian, Greek, Croatian, Serbian, Bosnian, Slovenian, Macedonian, Albanian, Danish, Swedish, Norwegian, Finnish, Estonian, Latvian, Lithuanian
Americas · 2
Spanish (Argentina), Portuguese (Brazil)
Middle East & Central Asia · 8
Arabic, Hebrew, Persian, Turkish, Azerbaijani, Armenian, Kazakh, Uzbek
South Asia · 8
Hindi, Bengali, Urdu, Gujarati, Marathi, Tamil, Telugu, Nepali
East & Southeast Asia · 12
Mandarin Chinese, Japanese, Korean, Mongolian, Vietnamese, Thai, Lao, Burmese, Indonesian, Javanese, Malay, Filipino
Africa · 1
Swahili
China region · 6
Mandarin (many voices), English, Cantonese, Sichuanese, Henan dialect, Hunan dialect
EVALUATION KIT
Hear it in your own room, in your own language, first
The basic voice mainboard, speaker and microphone can be bought on their own — the same voice front end our projects build on, flashed with dialogue firmware. Useful for measuring wake-up, recognition, response latency and sound before committing development budget.
Ask about the kit- Voice mainboard with dialogue firmware
- Speaker
- Microphone
- Bare boards, no enclosure or battery; runs from any USB-C charger — ships by regular express, no dangerous-goods paperwork
Typical projects
Four kinds of project we build
Anonymised archetypes from recent work — the configuration, and the problem that actually took the engineering time.
Clip-on in-car companion
Clamps to the head-unit display · offline wake word, direction finding and music-rhythm mode · head and body motors · Bluetooth / CAN / USB to the head unit · battery or 12 V
The hard bit
Road noise against the wake word, and a head that must turn to the driver or passenger without twitching at every horn.
Turntable desk companion
Spherical head on a motorised 360° base · RGB light ring · nod motor · optional camera · battery or USB-C
The hard bit
Rotating without end means power and signal across a moving joint — slip ring or a limited sweep — and keeping gear noise out of the mics.
Accessory-aware board-game plush
4G + GNSS · multiple NFC readers for accessories and tap-to-befriend · touch zones · motion sensing · moving ears · vibration · cloud-orchestrated play
The hard bit
NFC antennas that still read through fabric and stuffing, and a 4G battery budget that survives a full afternoon of play.
Bedside sleep companion
mmWave radar for presence and breathing · top-mounted PPG for heart rate and SpO₂ · LED matrix face · far-field mics · night light · Wi-Fi, optional 4G
The hard bit
Radar that does not mistake a fan or a pet for a sleeper, and health wording that stays on the right side of medical-device rules.
In depth
Request a Quote
Start a ai products project
Send the market, target price band and certifications. You get a platform recommendation, a cost envelope and the certification path — from an engineer, within one business day.