IPheromone

What if AI could connect you to the people you are meant to find — guiding you with synthetic pheromones?

Winner — Connect track, MIT Hard Mode: Hardware AI Hackathon (MIT Media Lab, 2026). Sponsors: Anthropic, Akamai, Qualcomm, Bambu Lab.

Built by Awu Chen, Paul-Peter Arslan, Jianing Yu, Richard Zhang, and Elisa over one weekend at MIT.

---

Demo

https://github.com/Polpii/IPheromone/raw/master/IPheromone_demo.mp4

If the player above does not render, download the demo video (3.9 MB).

---

The idea

We live in a world where technology keeps making us more disconnected — even when we are surrounded by people with overlapping interests, shared ambitions, and compatible histories. We never find each other.

IPheromone is a wearable ambient-intelligence system that changes that. It gets to know you deeply through a voice interview, distills your personality into a unique scent signature, and then quietly works on your behalf in social spaces — scanning for other devices nearby, running AI-to-AI compatibility conversations, and triggering a sensory signal when it finds someone you are highly compatible with.

No app to check. No swiping. Just a scent and a heartbeat.

---

Research framing — why this is interesting beyond the hackathon

IPheromone is built around three patterns that are central to modern LLM research:

1. User simulation with LLM personas. Each wearer has a personal LLM agent that holds a structured profile (built during the voice interview) and can answer questions as that user — a high-fidelity simulation of the wearer used for evaluation and for live agent-to-agent negotiation.
2. Multi-agent autorating. When two devices come within range, their agents exchange structured profile summaries and run a brief negotiation that ends with a calibrated compatibility score — i.e. an automatic rating of the human-to-human match, grounded in the user-simulation profiles above.
3. Closed-loop sensory feedback. A high score triggers a physical signal — scent release + an escalating haptic heartbeat — that uses the body as a low-bandwidth output channel for a high-stakes social inference.

This combination (user simulation → autorating → embodied feedback) is the same evaluation pattern used in LLM-judge research, just put into a wearable.

---

How it works

1. Interview. Your agent interviews you by voice. It asks about your personality, interests, history, and aspirations — then builds a rich profile using Claude (Anthropic).
2. Scent assignment. Your profile is distilled into a unique scent recipe — a sequence of fragrance channels (12 in the current diffuser) that represents who you are.
3. Social scanning. When you enter a space, your device scans for other IPheromone users nearby. Agents run 1-to-1 conversations with each, exchanging information and scoring compatibility.
4. Match signal. When a strong match is found, their scent is released around your neck via the wearable diffuser. A haptic heartbeat then begins, increasing in frequency as the two of you physically approach each other.

---

Tech overview

The system has three coordinated parts:

1. Web app (Next.js)

The control center. Handles user profiles, the interview flow, the agent-to-agent matching conversations, and the scent-recipe manager. Visible on phone or laptop in a shared space, and reachable on the local network from each Raspberry Pi.

2. Raspberry Pi wearable client

A physical device with a small display, a NeoPixel LED strip, a microphone, a speaker, and a push-to-talk button. It shows the companion's state in real time (a pixel-art Tamagotchi), lets the wearer speak by holding the button, and drives the sensory feedback loop (scent + heartbeat).

3. BLE scent bridge

A Python HTTP server running on the same machine as the web app. It exposes a small REST interface and translates calls into BLE commands sent to the scent-diffuser hardware. The diffuser exposes 12 fragrance channels that can be opened individually and on a schedule.

---

Features

• Voice interview — Claude conducts a deep onboarding conversation to build your profile.
• Unique scent recipe — AI maps your personality to a sequence of fragrance channels on the diffuser.
• Multi-agent matching — when two wearers are nearby, their AI agents have a direct conversation and score compatibility.
• Physical feedback — a match triggers scent release + an escalating heartbeat sequence (MP3).
• Tamagotchi companion — a pixel-art creature lives on the Pi display, animating in sync with the system state (listening, thinking, dating, etc.).
• LED strip animations — NeoPixels react to every state change in real time.
• Push-to-talk — hold the button on the Pi to talk; Whisper transcribes and the agent responds.

---

Project structure

├── src/
│   ├── app/
│   │   ├── page.tsx              # Main dashboard
│   │   ├── scent/                # Scent device control & recipe viewer
│   │   ├── user/[id]/            # Per-user companion screen
│   │   └── api/                  # API routes (see below)
│   ├── components/
│   │   ├── CompanionScreen.tsx   # Main interactive companion UI
│   │   ├── Tamagotchi.tsx        # Pixel-art creature renderer
│   │   ├── InterviewScreen.tsx   # Onboarding interview flow
│   │   └── WelcomeScreen.tsx
│   └── lib/                      # Shared utilities (Claude client, profiles, scent bridge)
├── raspberry/
│   ├── client.py                 # Pi hardware client
│   ├── start-client.sh           # Boot script
│   └── install-autostart.sh      # One-time setup on the Pi
├── scent-bridge/
│   └── server.py                 # BLE bridge (Python HTTP → BLE diffuser)
└── data/                         # JSON store (profiles, recipes, conversations, states)

---

Getting started

Requirements

• Node.js 18+
• An Anthropic API key (Claude)
• Python 3.10+ (for the scent bridge and the Pi client)

Install

npm install

Create .env.local:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...        # for Whisper audio transcription
SCENT_BRIDGE_PORT=5050       # optional, defaults to 5050

Run

npm run dev

Opens on http://localhost:3000 — and on your local network at http://<your-ip>:3000, which the Pi client connects to.

Keyboard shortcuts (web app)

Key	Action
L	Toggle voice listening
A	Show artwork on the Pi display for 30 s

---

Raspberry Pi setup

Hardware

Component	Role
Raspberry Pi 4	Main compute
Seeed Studio ST7789	240×280 SPI display
NeoPixel LED strip	30 px, GPIO13
Push button	GPIO26
USB microphone	Audio input
Speaker + mpg123	Audio output

Install dependencies on the Pi

sudo apt update && sudo apt install -y mpg123 python3-venv

python3 -m venv ~/display-env
source ~/display-env/bin/activate

pip install \
  adafruit-circuitpython-rgb-display \
  adafruit-circuitpython-neopixel \
  Pillow requests sounddevice numpy \
  gpiozero bleak

Autostart on boot (one-time setup)

Copy client.py, start-client.sh, and install-autostart.sh to /home/ipheromone/raspberry/, then:

sudo bash /home/ipheromone/raspberry/install-autostart.sh "YOUR_WIFI_PASSWORD"

After that, plug the Pi in and it connects to Wi-Fi and starts the client automatically — no interaction needed.

Run manually

source ~/display-env/bin/activate
cd ~/raspberry
python3 client.py --server http://<server-ip>:3000 --user elisa

Sound files (place in raspberry/)

File	Trigger
listen.mp3	Recording starts
wave.mp3	Someone waves
dating.mp3	Dating loop (plays until match found)
heartbeat05.mp3	Match found — slow heartbeat
heartbeat1.mp3	Approaching — medium heartbeat
heartbeat15.mp3	Close — fast heartbeat

---

Scent bridge

The bridge runs as a local Python HTTP server and translates REST calls from the web app into BLE commands sent to the scent diffuser. It auto-starts when the /scent page opens, but you can also run it manually:

cd scent-bridge
pip install bleak
python server.py --port 5050

Route	Method	Purpose
/health	GET	Status (no BLE scan)
/connect	GET	Scan and connect
/play	POST	Play a scent sequence
/stop	POST	Force stop all 12 channels

---

API routes

Route	Purpose
POST /api/interview	Claude voice interview
POST /api/listen	Process transcription, route to action
POST /api/transcribe	Audio → text (Whisper)
GET/POST /api/user/state	Tamagotchi state sync (Pi ↔ web)
GET/POST /api/profile	User profiles
POST /api/profile/register	Register new user
POST /api/seed	AI-generate a user profile
POST /api/agents/date	Run agent-to-agent matching
GET /api/agents/activity	Active matching sessions
GET /api/agents/conversations	Conversation history
POST /api/scent/play	Play scent recipe
POST /api/scent/stop	Stop playback
POST /api/scent/recipe	Generate scent recipe
GET /api/scent/status	Device status
POST /api/scent/connect	Connect to BLE device

---

Stack

Layer	Technology
Framework	Next.js 16 + React 19
Language	TypeScript
AI	Claude (Anthropic) — interviews, profiles, matching
Transcription	OpenAI Whisper
Styling	Tailwind CSS 4
Hardware AI	Raspberry Pi 4
Display	Adafruit CircuitPython ST7789
LEDs	NeoPixel
BLE	bleak (Python)

---

Recognition

🏆 Winner — Connect track, MIT Hard Mode: Hardware AI Hackathon, MIT Media Lab, 2026. Sponsors: Anthropic, Akamai Technologies, Qualcomm, Bambu Lab.

---

License

MIT. See LICENSE.