synapse

What if your phone could listen to your life, understand it through AI, and correlate what you said with how your body was feeling — in real time?

---

The Vision

Most personal knowledge tools (notes apps, journals, voice memos) require you to manually record, organize, and retrieve information. SYNAPSE removes that friction entirely.

You wear a Mi Band 5. Your phone listens. SYNAPSE captures ambient audio, transcribes it in real time via Google Gemini, tags it with biometric context (heart rate, activity state from the wearable), stores everything in a structured Supabase backend, and triggers automation workflows via n8n — all without you touching your phone.

The result: a searchable, contextualized personal knowledge base built passively from your own life.

---

System Architecture

┌────────────────────────────────────────────────────────────────┐
│                        ANDROID DEVICE                              │
│                                                                    │
│  Microphone ─► AudioCapture Service ─► Gemini API ─► Transcript      │
│                                                        +            │
│  Mi Band 5  ─► BLE GATT Client  ─────────► Biometric Snapshot  │
│              (HR + activity)                                        │
└────────────────────────────────────┬─────────────────────────┘
                                     │
              ┌────────────────────┴─────────────────┐
              │                   CLOUD                   │
              │                                           │
    ┌────────┴───────┐         ┌───────────────────┘
    │  Supabase (PostgreSQL) │         │  n8n Workflows         │
    │  sessions              │         │  ► Daily AI summary    │
    │  biometric_snapshots   │         │  ► WhatsApp delivery   │
    │  RLS auth policies     │         │  ► Archive raw audio   │
    └───────────────────────         └───────────────────┘

---

Core Features

🎤 Real-Time Audio Transcription

Background service captures audio in configurable chunks (2s–10s windows). Each chunk is sent to the Gemini API for transcription and results are streamed back and appended to the current session record, giving users a live transcript.

Why Gemini over Whisper or on-device STT? Gemini handles code, technical terms, and mixed-language input (PT/EN) without specialized fine-tuning. API latency is acceptable for non-conversational capture (≤1.5s per chunk at LTE). On-device STT accuracy drops significantly for technical vocabulary.

⌚ Wearable Biometric Sync (Mi Band 5 via BLE)

A GATT client connects to the Mi Band 5 over Bluetooth Low Energy and polls:

• Heart rate (characteristic UUID: 0x2A37)
• Activity state (resting / active / workout)

Biometric snapshots are timestamped and correlated with transcript records — enabling queries like "show me everything I said when my heart rate was above 90".

☁️ Structured Cloud Storage (Supabase)

CREATE TABLE sessions (
    id           UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id      UUID REFERENCES auth.users NOT NULL,
    started_at   TIMESTAMPTZ DEFAULT NOW(),
    ended_at     TIMESTAMPTZ,
    transcript   TEXT,
    duration_sec INTEGER
);

CREATE TABLE biometric_snapshots (
    id           SERIAL PRIMARY KEY,
    session_id   UUID REFERENCES sessions(id),
    recorded_at  TIMESTAMPTZ DEFAULT NOW(),
    heart_rate   SMALLINT,
    activity     VARCHAR(20)
);

Row Level Security policies ensure only the authenticated user can access their own records.

⚡ Automation Layer (n8n)

Webhook triggers fire when a session ends. n8n workflows generate a daily AI summary (Gemini prompt across all sessions from the last 24h) and deliver it via WhatsApp through Evolution API.

---

Tech Stack

Layer	Technology	Rationale
Mobile	Kotlin, Android (min API 26)	Coroutines + Flow for async audio pipeline
AI	Google Gemini Flash 1.5	Best multilingual STT quality without fine-tuning
BLE	Android Bluetooth LE (GATT)	Consumer wearable integration without official SDK
Backend	Supabase (PostgreSQL + Auth)	Zero-server backend; RLS for data isolation
Automation	n8n (self-hosted)	Webhook-driven pipelines; no vendor lock-in
Messaging	Evolution API	WhatsApp delivery of daily AI summaries

---

Engineering Challenges

Battery & Background Execution — Android kills background processes. SYNAPSE uses a Foreground Service with a persistent notification to keep audio capture alive. BLE polling uses WorkManager with battery-aware constraints.

BLE Reverse Engineering — The Mi Band 5 has no official SDK. HR characteristic UUIDs were mapped by inspecting GATT service tables manually. The connection retry logic handles the band's aggressive power-saving disconnects with exponential backoff.

Chunk Latency Budget — Each audio chunk must complete the full pipeline (capture → Gemini API → UI) within the next chunk's recording window to avoid transcript gaps. Current 3s windows give ~1.5s budget for API round-trip. Overflow chunks are queued and displayed in order on flush.

---

Roadmap

[v0.9  — current]  Core audio capture → Gemini transcription → Supabase storage
[v0.9.5]            Mi Band 5 BLE HR overlay on transcript timeline
[v1.0]              n8n daily summary → WhatsApp delivery (full pipeline live)
[v1.1]              Semantic search across transcripts (pgvector embeddings)
[v1.2]              Offline-first mode with cloud sync on reconnect
[v2.0]              Multi-device support + shared session spaces

---

Getting Started

Requires a Google Gemini API key and a Supabase project (both have free tiers).

git clone https://github.com/MarcosCF1/synapse.git
# Open in Android Studio (Hedgehog / Iguana+)
cp local.properties.example local.properties
# Set: GEMINI_API_KEY, SUPABASE_URL, SUPABASE_ANON_KEY
# Connect Android device (API 26+) → Run 'app'

BLE features require a physical device. Pair Mi Band 5 via system Bluetooth before launching the app.

---

Author

Marcos Pires — Full Stack Developer & AI Integration Engineer

Building at the intersection of AI, mobile, and automation. Open to Data Engineering and Backend roles.

---

Active development — star the repo to follow progress.