Skip to the content
2 layers28.8 × 41.5 mmLayouts, the printed circuit board in 3D

Vibe electronics AI pin

Size
28.8 × 41.5 mm
Copper layers
2
Thickness
1.6 mm
Parts
46
Nets
56
Pads
208
Tracks, vias
368, 49
Pours
8
Routed
Yes
KiCad
KiCad 9.0 (file 20241229)
Created
10 Aug 2025
Updated
5 Mar 2026

Overview

Layouts is an open-source printed circuit board by atopile, from the atopile/ai-pin repository on GitHub, published under the MIT licence. It is a 2-layer board of 28.8 × 41.5 mm with 46 parts on 56 nets, routed with 368 tracks and 49 vias.

Its parts include 18 resistors, 13 capacitors, 5 ICs, 3 diodes and 2 switches.

README

README.md

AI Pin — Your private, wearable memory

AI Pin front render

Stop forgetting what matters. AI Pin is a privacy‑first wearable that captures your conversations throughout the day, then—when you’re home—offloads to your own local server to transcribe and make everything searchable. Ask, “What did I promise Alex yesterday?” or “What did I get done today?” and get instant answers from your own life.

Why this is different

  • AI‑designed hardware: The PCB is designed using atopile (a declarative electronics DSL) and lives as code in ai-pin.ato.
  • Local‑first by default: All storage, transcription, embeddings and search run on your own machine—no cloud or third‑party services.
  • End‑to‑end system: Wearable capture → local ingestion → Whisper transcription → vector search → Q&A and daily summaries.

What it does

  • Continuous or on‑demand capture via an I2S MEMS mic, storing audio on microSD.
  • Nightly offload over Wi‑Fi while charging to a local server (Docker Compose).
  • Local transcription using Whisper (GPU if available, CPU fallback).
  • Search & recall across transcripts with semantic + keyword search.
  • Daily digest of your day’s conversations and key moments.

Moments you’ll stop forgetting

  • “What did I promise to follow up on in today’s meeting?”
  • “Who did I meet at the event—and what did we agree to do next?”
  • “Which book/podcast did Sam recommend at lunch?”
  • “What were the price, dates, and address the contractor mentioned?”
  • “Where did we leave off on that debugging thread last week?”
  • “What did the coach say about pickup changes for tomorrow?”
  • “What’s the Wi‑Fi password/door code they told me?”
  • “What exact phrasing did I use when I committed to that deliverable?”
  • “What were the three ideas I said I’d write down later?”

Architecture

mermaid
graph TD
  Wearable["ESP32‑C3 AI Pin<br/>I2S mic → microSD (WAV)"] -->|Wi‑Fi offload| API[FastAPI Ingestion]
  API --> MINIO[(MinIO Object Storage)]
  API --> REDIS[(Redis Queue)]
  REDIS --> WORKER[Whisper Worker]
  MINIO --> WORKER
  WORKER --> MINIO
  WORKER --> DB[(Postgres + pgvector)]
  UI[Search & Daily Digest UI] --> API
  API --> DB

Processing flow (upload → answer)

mermaid
sequenceDiagram
    participant Dev as Device
    participant API as Ingestion API
    participant S3 as MinIO
    participant Q as Redis Queue
    participant W as Whisper Worker
    participant DB as Postgres/pgvector

    Dev->>API: POST /v1/upload (WAV)
    API->>S3: putObject(wav)
    API->>Q: enqueue({ key, device_id })
    W-->>Q: blpop()
    Q-->>W: { key, device_id }
    W->>S3: getObject(wav)
    W->>W: transcribe (Whisper)
    W->>S3: putObject(txt/json)
    W->>DB: upsert chunks + embeddings
    Note over W,DB: hybrid index: BM25 + vector
    UI->>API: /v1/search?q=...&time=...
    API->>DB: retrieve + rank
    API-->>UI: results/snippets

Hardware (designed in code)

  • MCU: ESP32‑C3 Mini‑1 (I2S, SPI, Wi‑Fi)
  • Mic: TDK InvenSense ICS‑43434 (digital I2S)
  • Storage: microSD over SPI
  • Power: USB‑C → BQ25185 charger/power‑path → LDO (3.3 V)
  • Battery: LP402535 (configurable)

See ai-pin.ato for the complete, declarative hardware design.

Firmware (Zephyr RTOS)

  • I2S capture → ring buffer → WAV writer (5–10 min segments)
  • FATFS on SPI microSD, crash‑safe finalize, sidecar metadata
  • Wi‑Fi offload window while charging; TLS upload
  • Simple UX: LED states and a single button

Local backend (Docker Compose)

  • Ingestion API (FastAPI)
  • Object storage (MinIO)
  • Database (Postgres + pgvector)
  • Queue (Redis)
  • Transcription worker (Whisper via faster-whisper)

Agents and your private history

You can point a local agent at your private transcripts to answer questions and automate tasks—without sending data to the cloud.

  • Local‑only by default: Retrieval happens against your local DB/object store. No outbound calls unless you enable them.
  • Hybrid retrieval: BM25 + vector search with time filters and device/user scoping.
  • Defensive prompts: The API returns citations and timestamps; agents are encouraged to include them in answers.

Planned Agent API (local):

  • GET /v1/search – hybrid search over transcripts
  • POST /v1/agents/query – higher‑level query that returns stitched context windows for LLMs
  • POST /v1/agents/sessions – create a scoped session with time/person/topic constraints

Example (planned) search call:

bash
curl "http://localhost:8080/v1/search?q=what+did+I+promise+Alex&from=2025-01-01&to=2025-01-31&top_k=10"

Privacy boundaries:

  • Local network only; bind to loopback or LAN per your .env.
  • API tokens for agent clients; optional mTLS.
  • Audit log of queries; per‑user/device scoping.
  • Optional redaction layer (emails, credit cards) before indexing.

Quickstart

1) Start the local stack

Create backend/compose/.env with these defaults (edit as needed):

env
API_HOST=0.0.0.0
API_PORT=8080

MINIO_ENDPOINT=minio:9000
MINIO_ACCESS_KEY=minio
MINIO_SECRET_KEY=minio12345
MINIO_SECURE=false
MINIO_BUCKET=aipin

REDIS_URL=redis://redis:6379/0
WHISPER_MODEL=small
WHISPER_COMPUTE_TYPE=auto

Then launch:

bash
cd backend/compose
docker compose up -d

Services:

  • API: http://localhost:8080
  • MinIO console: http://localhost:9001 (user: minio, pass: minio12345)
  • Postgres: localhost:5432
  • Redis: localhost:6379

Health check:

bash
curl http://localhost:8080/healthz

2) Upload a sample WAV (simulating the device)

bash
curl -X POST \
  -F "device_id=test-device" \
  -F "file=@/path/to/sample.wav;type=audio/wav" \
  http://localhost:8080/v1/upload

The worker will pick it up, transcribe, and write a .txt next to the .wav in MinIO.


Developing the firmware (ESP32‑C3 + Zephyr)

See detailed instructions in firmware/zephyr/README.md. Quickstart:

Prereqs (one‑time):

bash
/opt/homebrew/bin/python3 -m venv "$HOME/zephyr-venv"
source "$HOME/zephyr-venv/bin/activate"
pip install -U pip west
cd "$HOME/zephyrproject"
export ZEPHYR_SDK_INSTALL_DIR="$HOME/zephyr-sdk-0.17.2"
export ZEPHYR_TOOLCHAIN_VARIANT=zephyr
west zephyr-export

Build & flash:

bash
source "$HOME/zephyr-venv/bin/activate"
cd "$HOME/zephyrproject"
west build -b esp32c3_devkitm $(pwd -P)/../projects/ai-pin/firmware/zephyr/app -p always -- \
  -DDTC_OVERLAY_FILE=boards/esp32c3_devkitm.overlay \
  -DPython3_EXECUTABLE="$HOME/zephyr-venv/bin/python" -DWEST_PYTHON="$HOME/zephyr-venv/bin/python"
west flash

Provision Wi‑Fi on SD card (root):

ini
ssid=YOUR_SSID
psk=YOUR_PASSWORD

Troubleshooting highlights:

  • Missing blobs for ESP32: west blobs fetch hal_espressif
  • I2S DMA error: ensure overlay enables &dma { status = "okay"; } and &i2s { dmas = <&dma 0>, <&dma 1>; }, and CONFIG_DMA=y in prj.conf
  • CMake can’t find west: pass venv python flags shown above

Project status

  • Hardware captured in ai-pin.ato (Atopile)
  • Backend Compose stack scaffolded: API + MinIO + Postgres + Redis + Whisper worker
  • Firmware stubs prepared (Zephyr app structure)

See spec.md and todo.md for detailed scope, milestones, and open questions.


Roadmap (MVP highlights)

  • End‑to‑end ingest → transcribe
  • Daily summary generation
  • OTA updates via MCUboot
  • Optional SD encryption and mTLS device identity

Contributing

Issues and PRs welcome. The goal is a great local‑first reference design for private, wearable memory.


License

MIT

Parts

46
QtyValueFootprintReferences
1310kΩ ±1% 62.5mWUNI_ROYAL_0402WGF1002TCE:R0402R1, R2, R3, R5, R6, R7, R11, R12, R14, R15, R16, R17, R19
7100nF ±10% 16V X7RSamsung_Electro_Mechanics_CL05B104KO5NNNC:C0402C3, C4, C6, C7, C8, C12, C13
41µF ±10% 25V X5RSamsung_Electro_Mechanics_CL05A105KA5NQNC:C0402C1, C2, C9, C11
2ALPSALPINE_SKRPACE010:KEY-SMD_4P-L4.2-W3.2-P2.20-LS4.6SW1, SW2
25.1kΩ ±1% 62.5mWUNI_ROYAL_0402WGF5101TCE:R0402R9, R10
1STMicroelectronics_LDK220M_R:SOT-23-5_L3.0-W1.7-P0.95-LS2.8-BLU1
1BQ25185DLHR_BQ25185DLHR:WSON-10_L2.2-W2.0-P0.40-BL-EP_TI_DLH0010AU5
1Korean_Hroparts_Elec_TF_01A:TF-SMD_TF-01ACard1
1MOLEX_532610271:CONN-SMD_532610271U4
1OPSCO_Optoelectronics_SK6805_EC20:LED-SMD_4P-L2.0-W2.0-BRLED1
1EEMB_BATTERY_LP402535:EEMB_BATTERY_LP402535B1
1Hubei_KENTO_Elec_KT_0603R:LED0603-RDLED3
1Espressif_Systems_ESP32_C3_MINI_1_N4:WIFIM-SMD_ESP32-C3-MINI-1U2
1TDK_InvenSense_ICS_43434:MIC-SMD_6P-L3.5-W2.7-P0.90-BLU3
1STMicroelectronics_USBLC6_2SC6:SOT-23-6_L2.9-W1.6-P0.95-LS2.8-BLD1
1BHFUSE_BSMD0805_050_15V:F0805F1
1SHOU_HAN_TYPE_C_16PIN_2MD_073:USB-C-SMD_TYPE-C-6PIN-2MD-073USB1
1atopile_Logo_4x4mm:atopile_Logo_4x4mmLOGO1
1100kΩ ±1% 62.5mWUNI_ROYAL_0402WGF1003TCE:R0402R8
110µF ±20% 6.3V X5RMurata_Electronics_GRM155R60J106ME44D:C0402C10
116kΩ ±1% 62.5mWUNI_ROYAL_0402WGF1602TCE:R0402R4
118kΩ ±1% 62.5mWYAGEO_RC0402FR_0718KL:R0402R13
122µF ±20% 6.3V X5RSamsung_Electro_Mechanics_CL10A226MQ8NRNC:C0603C5

Similar boards

  • 2 layers83.3 × 65.3 mmSimonShield, the printed circuit board in 3D

    SimonShield

    github-linguist/linguist

    Language Savant. If your repository's language is being reported incorrectly, send us a pull request!

    46 parts14kMIT
  • 2 layers27.8 × 22.3 mmAudio, the printed circuit board in 3D

    Audio

    byrantech/laptop

    anyon_e, a highly integrated, high end, open source laptop. Attempt the impossible.

    46 parts4.3kMIT
  • 2 layers370 × 11 mmDaylight XXL, the printed circuit board in 3D

    Daylight XXL

    VoronDesign/Voron-Hardware

    Cool Hardware designed by the Voron Design group

    46 parts1.1kGPL-3.0
  • 2 layers87.5 × 50 mmMesabob, the printed circuit board in 3D

    Mesabob

    rene-dev/stmbl

    AC Servo Driver for STM32F4

    46 parts990GPL-3.0
  • 2 layers38.8 × 62.5 mmESPAudio, the printed circuit board in 3D

    ESPAudio

    hackclub/OnBoard

    💡 Join 1,000 teenagers and make your first PCB with a $100 grant!

    46 parts915
  • 2 layers53.2 × 24.8 mmPicoLink, the printed circuit board in 3D

    PicoLink

    ErichStyger/mcuoneclipse

    McuOnEclipse projects and examples

    46 parts765
  • 2 layers23 × 29 mmESP32-epaper adapter, the printed circuit board in 3D

    ESP32-epaper adapter

    pcbreflux/espressif

    all espressif stuff will committed here

    46 parts608GPL-3.0
  • 2 layers25.5 × 73.5 mmTERES PCB2-IO Rev.B, the printed circuit board in 3D

    TERES PCB2-IO Rev.B

    OLIMEX/DIY-LAPTOP

    Do It Yourself Open Source Hardware and Software Modular Hacker's Friendly Laptop

    46 parts578Apache-2.0

Questions

How big is the Layouts PCB?

It measures 28.8 × 41.5 mm and has 2 copper layers, 1.6 mm thick.

What parts does the Layouts use?

46 parts. The full list of its parts is on this page, with links to each part's symbol and footprint where PCBKit's library has them.

Where can I download the Layouts KiCad files?

From the atopile/ai-pin repository on GitHub, at the commit this page was made from: the board is layouts/default/default.kicad_pcb. PCBKit opens the board in the browser, with nothing to install.

Can I use the Layouts design in my own project?

Its repository publishes it under the MIT licence: you may reuse it on that licence's terms (keep the author's notice, and for a copyleft licence share your changes alike).