🌌 Unmute1ai — Calm Tech Productivity Engine

"Not a productivity app. A nervous system recalibration."

Built for the NYC Self-Improving Agents Hackathon — a neurodivergent-first productivity engine that eliminates executive dysfunction through agentic AI, zero-shame design, and radical local privacy.

🧠 The Problem

Neurodivergent high-performers (ADHD, autism, dyslexia spectrum) experience executive dysfunction as a wall — not laziness. Existing tools punish non-completion with red badges, streaks, and guilt spirals. Unmute1ai is different: it meets users in the frozen moment and gently moves them forward.

🏗️ Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                      UNMUTE1AI SYSTEM                        │
│                                                             │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐  │
│  │   FRONTEND   │    │  AGENT LAYER │    │  DATA LAYER  │  │
│  │              │    │              │    │              │  │
│  │ React + Vite │◄──►│ DeepThink    │◄──►│ AES-256-GCM  │  │
│  │ Tailwind CSS │    │ (Gemini 3.1) │    │ Local Vault  │  │
│  │ WCAG 2.2 AAA │    │              │    │              │  │
│  └──────────────┘    │ Arctic-2     │    │ IndexedDB    │  │
│                      │ (Snowflake)  │    │ (encrypted)  │  │
│  ┌──────────────┐    │              │    └──────────────┘  │
│  │ OBSERVABILITY│    │ Weave Traces │                       │
│  │              │    │ (W&B)        │    ┌──────────────┐  │
│  │ W&B Dashboard│◄──►│              │    │ ACCESSIBILITY│  │
│  │ Self-Improve │    └──────────────┘    │              │  │
│  │ Feedback Loop│                        │ User1st U1AI │  │
│  └──────────────┘                        │ Real-time    │  │
│                                          │ WCAG 2.2 AAA │  │
│                                          └──────────────┘  │
└─────────────────────────────────────────────────────────────┘

🤖 Sponsor Integrations

1. Google DeepMind — Gemini 3.1 Pro (DeepThink Engine)

The heart of the "Unstick Me" feature. When a user is frozen, they press one button. Gemini 3.1 Pro's long-context reasoning decomposes their stated goal into exactly 3 Vibe Units — small, neurologically safe micro-tasks with estimated cognitive load scores.

Model: gemini-3.1-pro via Google AI Studio API
Location: src/agents/deepThink.ts
Prompt Strategy: Chain-of-thought decomposition with explicit cognitive load calibration
Output Schema: { vibeUnits: [{ id, title, estimatedMinutes, cognitiveLoad: 'low'|'medium' }] }

2. Snowflake — Arctic-2 (Silent Speech Synthesis)

Arctic-2 handles the "Whisper Input" mode — users can sub-vocalize or speak softly and the model transcribes with <50ms perceived latency via streaming chunks.

Integration: src/hooks/useSilentSpeech.ts
Architecture: WebSocket streaming to Snowflake Arctic-2 endpoint
Latency Target: <50ms first-token via chunked streaming
Fallback: Standard Web Speech API when offline

3. Weights & Biases — Weave Traces (Self-Improving Loop)

Every agentic reasoning step is wrapped in a W&B Weave trace. A nightly job analyzes friction points (tasks abandoned, "Unstick Me" re-presses) and automatically refines the DeepThink system prompt.

Integration: src/lib/wandb/weaveWrapper.ts
Traces: Task decomposition quality, completion rates, friction events
Self-Improve Script: scripts/selfImprove.ts — runs nightly, updates prompts via W&B Artifacts
Dashboard: Real-time metrics at your W&B project URL

4. User1st (U1AI) — Real-time WCAG 2.2 AAA Accessibility

Every component rendered passes through the U1AI accessibility audit hook. Non-compliant renders are blocked with a developer warning — accessibility is not optional, it's a build gate.

Integration: src/hooks/useAccessibilityAudit.ts
Standard: WCAG 2.2 AAA (strictest available)
Audit Timing: Post-render, pre-paint via useLayoutEffect
CI Gate: scripts/a11yAudit.ts fails the build if any component scores below AAA

🎯 Core Workflows

"Unstick Me" — Central Action

User clicks "Unstick Me"
  → Modal prompts: "What's the one thing you're avoiding?"
  → Gemini 3.1 DeepThink decomposes goal
  → 3 Vibe Units appear (max 25 min each)
  → W&B Weave logs the decomposition
  → User picks ONE to start
  → Timer starts (optional, non-alarming)

"Triage Mode" — Noise Elimination Dashboard

Simulated Jira + Slack feeds ingested
  → Arctic-2 classifies each item: Signal | Noise | Deferred
  → 80% noise filtered to "Later" pile
  → 20% signal surfaced in calm, priority order
  → Zero red colors, zero urgency theater

"Self-Improving Loop" — Nightly Refinement

W&B Weave collects: completion rates, re-press events, session duration
  → Friction score calculated per decomposition pattern
  → Gemini 3.1 re-ranks decomposition strategies
  → New system prompt committed to W&B Artifacts
  → Next day: improved DeepThink quality

📁 Directory Structure

unmute1ai/
├── src/
│   ├── components/
│   │   ├── Dashboard.tsx          # Triage Mode central view
│   │   ├── UnstickButton.tsx      # "Unstick Me" central action
│   │   ├── VibeUnitCard.tsx       # Individual micro-task display
│   │   ├── TriageItem.tsx         # Signal/Noise classified item
│   │   └── AccessibilityWrapper.tsx  # U1AI audit wrapper
│   ├── hooks/
│   │   ├── useDeepThink.ts        # Gemini 3.1 DeepThink hook
│   │   ├── useSilentSpeech.ts     # Arctic-2 speech hook
│   │   ├── useWeaveTrace.ts       # W&B Weave tracing hook
│   │   └── useAccessibilityAudit.ts  # U1AI real-time audit hook
│   ├── lib/
│   │   └── crypto/
│   │       ├── vault.ts           # AES-256-GCM encryption vault
│   │       ├── keyDerivation.ts   # PBKDF2 key derivation
│   │       └── types.ts           # Encrypted data types
│   ├── agents/
│   │   ├── deepThink.ts           # Gemini 3.1 agent logic
│   │   └── triageAgent.ts         # Jira/Slack triage agent
│   └── styles/
│       └── globals.css            # Deep Space theme variables
├── scripts/
│   ├── selfImprove.ts             # Nightly W&B prompt refinement
│   └── a11yAudit.ts               # CI accessibility gate
├── .clauderules                   # No-Shame UI enforcement rules
├── .env.example                   # Required environment variables
├── vite.config.ts
├── tailwind.config.ts
└── README.md

🚀 Quick Start

# Clone and install
git clone https://github.com/unmute1ai/core
cd unmute1ai
npm install

# Configure environment
cp .env.example .env
# Fill in: GEMINI_API_KEY, SNOWFLAKE_ARCTIC_ENDPOINT, WANDB_API_KEY, U1AI_KEY

# Run development server
npm run dev

# Run accessibility audit (CI gate)
npm run a11y

# Run self-improve loop manually
npm run self-improve

🎨 Design Philosophy — No-Shame UI

The .clauderules file enforces these principles at the component level:

❌ Never	✅ Always
Red badges or colors	Soft lilac (#B794F4) accents
Streak counters	Session presence (no judgment)
"You missed X days"	"Welcome back, ready when you are"
Alarming timers	Gentle pulsing progress
Dense information	Generous whitespace
Notifications	Opt-in gentle nudges only

Color System:

Primary Background: #0B0E14 (Deep Space)
Accent: #B794F4 (Soft Lilac)
Surface: #141820
Text Primary: #E8EAF0
Text Muted: #6B7280
Success: #68D391 (soft green — never red)

🔐 Security Model

All user-generated content is encrypted before it leaves the browser:

Key Derivation: PBKDF2 with 310,000 iterations (NIST recommended)
Encryption: AES-256-GCM with random 96-bit IVs per write
Storage: Encrypted ciphertext in IndexedDB — never plaintext on disk
Key Storage: Derived key in memory only — cleared on session end
Zero Server-Side: No user content ever transmitted to any server

📊 W&B Self-Improvement Metrics

Metric	Description
`decomp_quality`	% of Vibe Units completed without re-decompose
`friction_score`	Re-press rate on UnstickMe button
`triage_accuracy`	User override rate on noise classification
`session_depth`	Tasks completed per session
`cognitive_load_calibration`	Actual vs. estimated completion time

🧪 Accessibility Compliance

Every component is tested against WCAG 2.2 AAA at render time via User1st U1AI:

Color Contrast: 7:1 minimum (AAA) — enforced by build gate
Focus Indicators: 3px solid visible focus rings
Motion: prefers-reduced-motion respected globally
Screen Readers: Full ARIA label coverage
Keyboard Navigation: 100% keyboard operable, no mouse traps

🏆 Hackathon Context

Event: NYC Self-Improving Agents Hackathon
Track: Calm Tech / Neurodivergent Accessibility
Team: Unmute1ai
Sponsors Used: Google DeepMind · Snowflake · Weights & Biases · User1st

Built with care for every brain that's ever felt stuck.

Built With

gemini
google-ai-studio
modulate
python

Updates

Matthew Sacks started this project — Feb 09, 2026 09:22 PM EST

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.