Our own AI for kids' watches: from buying the hardware to the hero's voice
AI system A local AI cluster in Almaty, models in three languages and the voice path of a kids' watch: the child asks, the hero answers in its own voice, and the child's voice and words never reach a cloud model.
Overview
OIYNUP is a learning universe for children from two to ten: an animated series, a game and a kids' watch with a hero living in it. I lead design and AI in the project: I assembled the company's own AI cluster, chose and tested the models and built the watch's voice path, from the child's question to the hero's answer in its own voice.
Hardware
In spring we went for the hardware ourselves, to Shenzhen with a stop in Dubai, and in August to China again. Every part was checked on the spot, part numbers, seals, stress tests, and only then paid for. Parallel research agents worked out the memory size and the architecture, and the final specification was later checked against the invoices.
Which machine does what
- Workstation
One card serves the watch's voice: recognition, the language model, synthesis. The other makes images, video and music for the team.
Threadripper PRO, 2 × RTX 5090, 256 GB ECC
- DGX Spark
Fine-tuning: a queue of training runs with a log, for example our own Kazakh voice.
NVIDIA GB10, 128 GB unified memory
- Network and power
A 10-gigabit switch, a direct 200-gigabit link between the two Sparks and an online UPS.
10 Gbit/s, 200 Gbit/s between the Sparks
The second DGX Spark now runs the local model of my personal system, the Vega case.
Models
The language model is ISSAI's Kazakh Qwen3.5-35B under an open licence: of the candidates, only it declined Kazakh words reliably and understood a picture from the camera. GigaAM recognises Russian, Whisper recognises Kazakh and English, Qwen3-TTS gives the hero its Russian voice. Every model was chosen by measurement, not by its description.
How a question becomes an answer
- Watch
The child presses the button and asks out loud.
KT38, Android
- Recognition
The language is detected, the speech becomes text.
GigaAM · Whisper
- Language model
Answers in two or three short sentences, as the hero would.
Qwen3.5-35B-A3B-Kazakh
- Safety
Dangerous topics get fixed answers, never improvised ones.
code around the model
- Voice
The answer comes back to the watch in the hero's voice.
Qwen3-TTS · CosyVoice3
All of it runs on the workstation in Almaty: no cloud model in the path.
Live run
On 5 November an adult asked a real KT38 watch why leaves turn yellow in autumn. Tria answered in her own voice in under two seconds: there is less sun in autumn, the green chlorophyll breaks down, and the yellow colours show.
Watch screens7
Safety is code around the model
The model does not decide on its own what is fine for a child. Code stands around it: fixed answers on dangerous topics, a check of the camera frame (the hero does not identify medicines, mushrooms or sharp objects and asks the child to call an adult) and an end-to-end set of child scenarios that passes in full. For now the team tests the watch; children get it only after the safety and legal review.
How we work
AI agents write the code under my direction: I set the task, Codex writes, another model reviews. An outside review took the first architecture draft apart into 26 remarks, 13 of them blocking, and all were accepted before work began. We load-tested the stack with synthetic requests without a single child, wrote down the weak spots, such as Kazakh recognition gluing words together, and are fixing them.
The OIYNUP agent
Next comes the team agent on the same local model. It answers employees about their project, keeps tasks and daily summaries and helps the engineer with model training. Rights follow data labels set by tools, not by the model: a person sees only their own project, children's data is closed to the agent entirely, and access, releases and money wait for the head's approval. Launching in 2026.
- Employee
Asks about their project or gives a task.
chat or Forgejo
- Data labels
The tool labels the request before the model sees it.
project · level · personal · children
- Local model
Answers, keeps tasks, writes summaries.
DGX Spark
- Approval
Anything irreversible waits for the head of the company.
access · release · money
Children's data lives in a separate process without internet; the agent cannot reach it.
In numbers
- 0cloud models in the pathA child's voice and words never leave our own hardware.
- 3languagesRussian, Kazakh and English: recognition, model and voice.
- 1.6 sto the hero's first wordLive on a real watch on 5 November 2026; 0.8 s on the test bench.
- 940test requestsA synthetic load run against the live stack, without a single child.
- 37child-safety scenarios, all passingEnd to end through the whole chain; multi-turn ones pass 8 of 8.
Credits
- AI and design: cluster, models, voice path, watch interface
- Jean Ivanov
- Code: AI agents under Jean’s direction, one writes and another reviews
- Codex, Claude
- Language model
- ISSAI, Qwen3.5-35B-A3B-Kazakh (Apache-2.0)
- Animated series «Мир Форм» and the heroes
- Grinvich Technology
Next projectA personal AI system: one agent and one app on every screenVega
