SafeHaven: Breaking the Silence on GBV

Inspiration

In Kenya, survivors of Gender-Based Violence (GBV) often face three massive barriers: Stigma, Cost, and Language. Many existing helplines only operate in English or during business hours, leaving millions vulnerable. I built SafeHaven to break the silence a digital sanctuary that is anonymous, free, and speaks the user's native language (Swahili or English), available 24/7.

What it does

SafeHaven is a multimodal AI crisis agent designed for immediate grounding and support.

• Conversational Voice Interface: Powered by ElevenLabs, the agent provides a lifelike, real-time voice experience that understands and speaks fluent Kenyan Swahili and English.
• PWA & Mobile Ready: Built as a Progressive Web App (PWA) with a fully responsive mobile-first design. Users can install SafeHaven directly to their home screens for discreet, instant access without needing to visit an app store.
• Safety First: Includes a "Whisper Mode" for discreet reporting in unsafe environments and a "Panic Button" that instantly scrubs the screen and redirects the browser.
• Grounding Tools: Features an offline "Leaf" tool that plays calming rain/ocean sounds to help users manage panic attacks in real-time.

How I built it

I built SafeHaven using a Next.js frontend to ensure high performance and seamless PWA capabilities.

• The Brain (Google Gemini 2.5 Flash): I leveraged Gemini 2.5 Flash for its incredible speed and multimodal capabilities. It acts as the core reasoning engine, performing complex safety analysis to detect high-risk emergencies vs. general stress. Its large context window allowed me to feed it specific cultural nuances of Sheng and Kenyan social contexts, ensuring the AI doesn't just translate, but truly "understands."
• The Voice (ElevenLabs Conversational Agent): I moved beyond simple Text-to-Speech by implementing the ElevenLabs Conversational Agent. This provides a full-duplex, low-latency conversational flow. By using the Turbo v2.5 model, I achieved the human warmth necessary for a counselor persona.
• Audio Intelligence: I integrated the ElevenLabs Voice Isolation API. This is crucial for our "Whisper Mode," as it effectively cleans background noise and enhances quiet speech before the audio reaches the model, ensuring accuracy even in high-stress, muffled environments.
• Responsive Architecture: Using Tailwind CSS and Next.js, I ensured the interface is fully responsive, catering to the wide variety of mobile devices used across Kenya, ensuring no user is left behind due to their hardware.

Challenges I ran into

• API Tier Constraints: Implementing a real-time crisis agent requires high reliability. I quickly hit the limits of the Gemini API free tier, where rate limiting threatened the "instant" nature of the support. Similarly, the advanced features of the ElevenLabs Conversational Agent required an upgrade to a Pro plan to handle the high-quality voice synthesis and concurrency needed for a production-ready tool.
• Latency vs. Empathy: Balancing the need for instant responses with the need for a warm, natural voice was tough. Using the combination of Gemini's Flash model and ElevenLabs' Turbo models was the key to reducing "dead air" that could otherwise increase a user's anxiety.
• Language Nuances: Standard AI models often struggle with "Sheng" (Kenyan slang). I had to carefully engineer the system prompts to ensure Gemini recognized phrases like "sina amani" or "nimechochwa" as indicators of psychological distress rather than literal statements.

Accomplishments that I'm proud of

I am most proud of the "Karaoke" Visual Alignment. Realizing that trauma survivors often experience "brain fog," I built a visualizer that highlights words in Teal as the ElevenLabs Agent speaks them. This visual-audio synchronization dramatically improves accessibility and helps users stay "grounded" in the conversation during a panic attack.

What I learned

I learned that Context is King. A generic AI is helpful, but a culturally aware AI is a lifeline. By fine-tuning the persona and leveraging the ElevenLabs ecosystem, I turned a standard chatbot into a high-empathy counselor. I also learned that in crisis tech, the "Conversation" is more than just words, it's the tone, the latency, and the ability to be heard through noise.

What's next for SafeHaven

I plan to introduce "Safe Zones" using geolocation to anonymously point users to the nearest verified physical shelter or hospital. I also intend to utilize Gemini’s Multimodal Vision capabilities, allowing users to share photos of legal or medical documents for the AI to summarize and explain in simple, comforting Swahili.

Built With

• antigravity
• elevenlabs
• gemini
• javascript
• next-js
• typescript