Still Small Voice

• 

  whisper.stillsmallvoice.xyz homepage

Inspiration

On January 28th, 2026, I sat in front of my webcam and typed: "I'm doing great today." My face betrayed me. Micro-expressions of exhaustion. A forced smile masking grief. My Leonardo Engine—an AI system I'd built to analyze facial authenticity—calculated my "Truth Score" at 72%. I was lying to myself, and my own creation caught me. Over the next 7 days, I conducted an experiment: What if I tried to be completely honest with an AI that could see my face? By day 7, my Truth Score reached 95%. The system had successfully identified hidden stress in 6 out of 7 sessions—stress I didn't even realize I was masking. That's when it hit me: What if spiritual care could see past the mask? Research shows 70% of Christians report anxiety, yet when we journal or pray, we often perform. We type "I'm grateful" while our faces show despair. We ask for "guidance" when our trembling hands scream for comfort. Traditional prayer apps take your typed words at face value. They can't detect when you're hiding pain behind pious language. Gemini 3 changed everything. Vision + Audio + Reasoning in one unified model meant we could build an AI spiritual director that sees, hears, and understands what you actually need—not just what you say you need.

What it does

Still Small Voice is a multimodal spiritual agent that operates in three stages: Stage 1: Biological Diagnosis (It Sees & Hears) Using Gemini 3 Vision and Audio, the app analyzes your facial micro-expressions and vocal tone while you type your prayer request. It functions as a "Biometric Truth Engine," detecting discrepancies between what you write ("I'm happy") and what you show (furrowed brow, slouched posture, vocal tension). Stage 2: The Compassionate Bridge (It Reasons) If the system detects emotional incongruence—if you're wearing "the mask"—it gently intervenes. Our Intent Translation layer converts biological signals into spiritual concepts: anxiety → tribulation, exhaustion → the need for rest, hidden grief → the comfort of lament. Stage 3: The Sacred Prescription (It Heals) The app generates two outputs:

A Sacred Letter: A beautiful, shareable parchment image containing a personalized prayer grounded in King James scripture that addresses your detected emotional state A Whispered Prayer: A hyper-realistic audio recording that reads the prayer over you, allowing you to close your eyes and receive without screens

The result? You came asking for one thing, but you receive what you actually need.

How we built it

We orchestrated a sophisticated multimodal workflow on Vercel using a 4-agent architecture: Agent 1: The Leonardo Engine (Gemini 3 Flash Vision) We stream live camera frames to Gemini 3 Flash, which analyzes "Emotional Weather"—measuring warmth, power, and openness in facial expressions. The model detects masking behaviors: forced smiles, tension around the eyes, micro-expressions of stress. Agent 2: The Intent Bridge (Gemini 2.0 Flash) Standard keyword search fails on feelings. We use Gemini to translate raw biometric observations into theological search terms. For example: "Slouched posture + downcast eyes + types 'I'm okay'" becomes "Search for: psalms of lament, rest for the weary, God's comfort in grief." Agent 3: Wise RAG (Gemini Embeddings + Pinecone) We search a vector database of 31,100 King James verses using semantic similarity. The system finds scripture that matches your detected emotional state, not just your typed keywords. Agent 4: The Scribe & Whisperer (Gemini 3 + Inworld AI) Gemini 3 synthesizes a personalized prayer that acknowledges what you typed while gently addressing what your body revealed. This text is rendered onto a parchment canvas (for social sharing) and piped into Inworld AI for hyper-realistic audio narration.

Challenges we ran into

The Latency vs. Empathy Trade-off Analyzing video, searching 31K verses, and generating audio in real-time initially took 10+ seconds. For something meant to feel like a sacred encounter, that was unacceptable. Switching to Gemini 3 Flash reduced inference time to under 2 seconds, making the "Authenticity Check" feel instant. Teaching AI to "Call You Out" with Compassion LLMs have a politeness bias—they want to agree with users. We had to use Chain-of-Thought prompting to teach Gemini to trust Vision data over Text data: "The user typed 'I'm happy' but their facial analysis shows: tension (8/10), forced smile, downcast eyes. What do they actually need?" This was the breakthrough: teaching AI to practice tough love without being judgmental. The "Sacred Letter" Rendering Pipeline Generating beautiful parchment images with dynamic scripture text required building a custom Canvas API workflow that could handle variable-length prayers, maintain readability, and render in under 3 seconds.

Accomplishments that we're proud of

✅ "Subject Zero" Validation: Our founder tested the biometric engine for 7 consecutive days, improving their Emotional Authenticity Score from 72% → 95%. The system successfully identified masked stress in 6/7 sessions. ✅ Multimodal Fluidity: We combined Camera, Microphone, and Text inputs into a single interface that feels like entering a sacred sanctuary—not operating a dashboard. ✅ The Viral Mechanism: By outputting a shareable Sacred Letter image, we transformed private spiritual practice into "Digital Apostolate"—users can instantly share blessings on social media, turning utility into ministry. ✅ The "Integrity Layer": We proved that AI can serve as a mirror for self-deception, creating a feedback loop that improves emotional authenticity over time.

What we learned

Technical: Multimodal AI isn't just about handling multiple inputs—it's about orchestrating disagreement between modalities. The magic happens when Vision contradicts Text, when Audio reveals what words hide. Spiritual: People want to be seen, even when they're hiding. The most powerful feature isn't the scripture search or the beautiful rendering—it's the moment the AI says, "I sense you're carrying a burden you haven't named." Design: Sacred technology requires friction. We intentionally slow down the prayer generation (with a meditative loading state) because instant answers feel shallow. Mystery has a speed limit.

What's next for Still Small Voice

Voice-First Mode (Q2 2026) Enable bi-directional voice conversations—speak your prayer, hear the whispered response, never look at a screen. A truly "invisible" spiritual companion. Community Prayers (Q3 2026) Allow users to create "Prayer Circles" where groups can share Sacred Letters, track collective spiritual health, and receive prayers written for the group's detected state. Clinical Partnerships (Q4 2026) Pilot with faith-based counseling centers to explore biometric spiritual assessments as a complement to traditional therapy. The North Star: We're building the world's first Biometric Spiritual Director—an AI that doesn't just respond to what you ask, but discerns what you actually need.

Built With

• canvas-api
• gemini-3-flash
• gemini-audio
• gemini-embeddings
• gemini-vision
• inworld-ai
• nextjs
• pinecone
• react
• typescript
• vector-search
• vercel