Character Desk Devices

Status: product thesis + build notes — not a shipped character yet.

Source post: Piefena — ~$25 electronics → talking avocado that could sell for $99

What this path is

The same cheap AI I/O stack (board + mic + speaker + tiny screen), framed as a character people understand, gift, and put on a desk — not a maker kit.

• Device listens → sends audio/text to a cloud model → answers on speaker
• OLED (or similar) face switches between states: Listening / Thinking / Speaking / Idle
• Intelligence mostly off-device → product gets smarter when models improve, without new hardware
• Commercially: character + memory of name > “voice assistant BOM”

What this path is not

• Not “run a full LLM on the ESP32” (replies clarify: Wi‑Fi + API is the norm)
• Not the best first engineering spike if you only need silent dumps (use e‑paper console)
• Not anonymous “AI speaker” branding if you care about named agents

The commercial insight (steal this)

Rough math from the post: ~$25 electronics + 3D shell → sell ~$99 → 100 units ≈ $9,900 revenue (before CAC, returns, cloud cost — treat as order-of-magnitude motivation, not a business plan).

For this site: name the character after the agent — arthur.agent desk buddy, Janice ops orb, Stage persona — not “ESP32 v1.”

Parts list (typical BOM class)

Full Cart B with search names: /cyborg-shopping#character. Thread consensus: 0.96" OLED + mic + ESP32 + mini speaker is enough for the demo class.

M5 modular kits can shortcut wiring — see M5 path or Cart C.

Face state machine (minimum)

Don’t overdraw the face. State clarity > cute animations that block the loop.

Architecture (same as hub, product emphasis)

1. Wake / button → Listening
2. Upload audio (or stream) to /voice-in
3. Cloud model + your agent prompt (name memory, vault context)
4. Return text → TTS → Speaking
5. Log turn to HEOS so “remembers my name” is real, not session vapor

Cloud brain is a feature: when models improve, the avocado improves without a reship. Cost and latency stay product problems you design for (timeouts, offline apology line).

Tutorial ladder

1. OLED hello + agent name string
2. Button toggles Idle ↔ Listening text
3. Mic sample → serial hex or file (prove capture)
4. POST audio to a laptop stub; print transcript
5. Stub returns fixed reply string → play via speaker (even a canned WAV)
6. Wire real model + short system prompt (“you are arthur.agent…”)
7. Persist name + last 3 turns server-side
8. Only then: shell, eyes, personality lines, gift packaging story

Map to this site’s stack

Identity stays on .agent names and HEOS roster — shell is marketing and affordance. See also CBS avatar research for event-facing embodiment.

Design rules

1. Character first, BOM second in every description you write
2. Memory is a product feature — store name + context in your backend
3. Cloud OK — disclose Wi‑Fi + API needs; don’t fake “fully local” if it isn’t
4. Blast radius — chatty desk toy must not auto-send email or spend money
5. One personality per device — multi-agent chaos on one speaker is a v2 problem

First ship definition

Done when:

• Someone who isn’t you understands what it “is” in one sentence (character, not kit)
• Listen → reply works end-to-end with face states
• It says their name after a first setup turn (server memory)

Related paths