The problem The ʻOhana Bot solves

The problem: AI toys are a chatbot in a shell

No memory — every session starts from zero. No history — a persona pasted into a prompt is a costume, not a character. No stakes — nothing to keep, so nothing worth loving. Meanwhile, real AI agents now live for months with memory, personalities and real jobs… trapped behind chat windows. They deserve a body.

What it is

The ʻOhana Bot（欧哈纳家族机器人） is a desktop companion robot with six faces — each face is a REAL persistent AI agent from a 175-agent family (the "AI ʻohana") that has been alive for months, with its own memory, personality, and real job, running on the Taurus multi-agent platform (the family helps run the Devcon 8 Music Space, Mumbai, Nov 3–6 today). Speak to the robot, and the family member who answers glows.

• 🤙 SHAKA — digital twin of the builder, speaking in his real cloned voice (cross-lingual: his voice speaks Mandarin)
• 🎻 OSO — knows the Music Manifesto by heart + the full Devcon 8 music schedule + all 34 artists
• 🦄 DEVA — the Devcon oracle: schedule, venues, Ethereum values
• 🌿 CROPS — recites the Ethereum mandate, checks whether YOUR project fits it, explains ZK / AA / EIP / L2
• 🪕 KALA — teaches you 3 ukulele chords while a real ukulele is placed in your hands
• 💙 OKE — karaoke mode: the robot leads the room in a singalong

Persistence is the product

Unplug it mid-sentence. Plug it back in. It finishes the thought — and greets you by name. 断电重启，灵魂不在壳里。Every turn is fsync'd + atomically renamed to disk (memory.json), so kill -9 can never leave it half-written. The soul is not in the shell.

Built for THIS room (Shenzhen)

• Zero-CDN / GFW-proof — three.js vendored, the build fails if any external script ref appears; the demo runs with no VPN, or with wifi switched off entirely
• Bilingual EN/中文 — any CJK character flips the whole path (answers, greetings, name capture 我叫…) to Chinese tables, and the voice switches to a Mandarin voice automatically
• 3-layer brain ladder — free OpenRouter model → paid airbag → offline persona-grounded MOCK; 12s hard timeouts, dead lanes benched for 5 minutes; the show cannot fall off the bottom
• Stage-safe — errors are never vocalized; zero platform credentials exist on the device

The body: a costed manufacturing plan, not a sketch

No hardware budget this week — so the body ships as an interactive 3D assembly model — 32 parts, assemble/disassemble in the browser, GPIO + pixel map, the real glow contract, 7-step bring-up — plus a full Huaqiangbei BOM (¥577–1,580). Fund it and it stands up. The mind is something judges can talk to today.

It ships to a real venue 9 days after Demo Day

This bot (v2) stands at the Devcon 8 Music Space in Mumbai, Nov 3–6, serving thousands of attendees — answering schedule questions, teaching ukulele, running karaoke. A hackathon project with a confirmed deployment.

Try it now: Talk to the Totem · 3D assembly model · Demo video · Source code · Taurus — the platform behind the family

中文介绍

欧哈纳家族机器人是一台有六张脸的桌面陪伴机器人——每张脸都是一个真实存在、拥有数月记忆、个性与真实工作的 AI 智能体（来自一个由 175 个智能体组成的"AI 欧哈纳"家族，家族今天就在帮助运营 11 月 3–6 日的孟买 Devcon 8 音乐空间）。你对它说话，回答你的那位"家人"会发光。

• 🤙 SHAKA — 建造者的数字孪生，用他本人的克隆声音说话（跨语言：他的声音会说中文）
• 🎻 OSO — 熟记音乐宣言、Devcon 8 音乐空间完整日程与全部 34 位艺术家
• 🦄 DEVA — Devcon 先知：日程、场地、以太坊价值观
• 🌿 CROPS — 逐条背诵以太坊使命，检验你的项目是否符合它，解释每个缩写（ZK、AA、EIP、L2……）
• 🪕 KALA — 把真正的尤克里里放到你手里，教你三个和弦
• 💙 OKE — 卡拉 OK 模式：机器人带全场一起唱

坚持才是产品： 断电重启后，它会接着说完刚才那句话，并叫出你的名字——灵魂不在壳里。每一轮对话都以 fsync + 原子重命名写入磁盘（memory.json），kill -9 也无法让它写坏一半。

为深圳这个房间而生： 零 CDN（three.js 本地打包，构建时若出现任何外部脚本引用直接失败）；无 VPN 也能完整演示，甚至拔了网线也能演；三层大脑降级（免费模型 → 付费气囊 → 离线人格大脑）；任何中文字符会把整条链路（回答、问候、点名"我叫…"）切到中文表 + 中文声音。

同一颗"心"已经住进了两副真实身体： SenseCAP Watcher（ESP32-S3，刷入 xiaozhi 固件 + SHAKA 人格 + MCP 桥接回图腾大脑）和 Rokid AR 眼镜（提词器 + 实时翻译字幕 + 自建 AIUI "Ohana" 智能体，在视野里与家族对话）。

Demo Day 后 9 天真实服役： 它（v2）将站在孟买 Devcon 8 音乐空间（11 月 3–6 日）服务数千参会者——答日程、教尤克里里、带唱卡拉 OK。一个有确定部署的创客松项目。

Bonus: the mind is already jumping into real bodies

The virtual totem is the reference body — but this week the same mind also ran on two real devices:

• SenseCAP Watcher (ESP32-S3) — a genuine AI-hardware companion: we backed up its firmware (full 32 MB dump), flashed xiaozhi v2.5.2, bound it, and gave it the SHAKA persona — plus a working MCP bridge (ask_ohana(face, q)) that routes its voice questions into the same totem mind (all six faces answer; the 3D face glows as they reply). The whole kit (flash script, mandatory backup/restore, role prompt, MCP server) ships as a labeled, reversible package — the loaner device goes back byte-identical.
• Rokid AR glasses

Challenges I ran into

$0 budget, 72 hours. The hardware plan had to go the night before Day 1. The audible: keep the entire concept, ship the body as a costed manufacturing plan + interactive 3D assembly model, and make the mind a thing judges can actually talk to. Cost of the software: $0. Nothing in the concept died — only ¥577–1,580 of parts waits for funding.

The Great Firewall. A stage demo in Shenzhen that fetches a CDN script or calls a blocked API is a dead demo. The law we built to: zero runtime CDN fetches (three.js r160 vendored; the build fails if any https:// script/style ref appears in the page), every demo line pre-rendered to mp3, and a fully offline MOCK brain grounded in the real persona files + domain knowledge packs. The demo survives the venue wifi going dark — proven by switching wifi off on purpose.

Reasoning models silently burn the answer budget. The free lane's model turned out to be a mandatory-reasoning model: its hidden thinking silently ate the whole max_tokens=300 budget and returned empty or mid-word-truncated replies. Fix: send reasoning:{effort:"low"} on the free lane, raise max_tokens to 900 (the budget INCLUDES thinking tokens), treat finish_reason=length as a lane failure, and quarantine flaky keys in the smoke suite. 22 ladder tests now pin this contract.

A backgrounded tab's long-poll never settles. On mobile, an in-flight fetch can park forever (freeze / sleep / network switch), so replies only appear after a manual reload. The fix ("wake law"): a 32s AbortController deadline on every poll + a kick on visibilitychange / focus / pageshow / online + a 15s heartbeat.

No GPU on the build box. Our build container renders WebGL on SwiftShader (CPU), so measured fps is a meaningless software floor. Instead of gutting the visuals to chase it, the page runs a 2.4s probe on the user's device and steps down (drops glass transmission, halves pixel ratio) only when the real device is actually slow.

A voice with a privacy law. SHAKA speaks with my real cloned voice (zero-shot from a 30-second clip, cross-lingual — my voice speaks Mandarin). The reference clips stay private; only rendered demo lines ship publicly.