Hearth

Landing Page
Translation Page
Language Selection
Support Page
Prompts
transcript
Aggressive expression detection + blocking translation
Alert exmaple

Inspiration

Emily and Jarny work the front desk at The Bloom Group's women's shelter - a low-barrier women's shelter in Vancouver. When we called them with them, we came expecting to hear about housing waitlists and digital access. What we didn't expect was that the thing they wanted most - the thing Emily named immediately when we asked "what would make your job easier?" - was interpreters.

For women in shelters, communication is not just about convenience, it is directly tied to safety and access to care. Many women are navigating crisis situations such as domestic violence, housing instability, or displacement, often while managing emotional stress, trauma, or caring for children. In these moments, being able to clearly express needs and understand support options is critical.

However, many shelters serve women from diverse linguistic and cultural backgrounds, where language barriers make already difficult situations even more isolating. When communication breaks down, it affects everything from intake and safety planning to accessing resources and receiving emotional support.

Through our conversations with the staff at The Bloom Group, we identified a clear and urgent need for a more reliable and accessible communication environment.

Shelter workers shared that:

Language barriers are a daily challenge, especially with residents from diverse backgrounds, including refugees and immigrants
Existing tools like Google Translate are not reliable enough in practice
Some languages, particularly African and Middle Eastern languages, are poorly supported
They don’t have the resources to hire interpreters
And sometimes, they don’t even know what language a resident is speaking

Beyond our interviews, data reinforced the scale of the issue:

Around 45% of shelter residents are non-native English speakers
About 87% have low to medium digital literacy

This means most existing solutions simply don’t work in this environment. And when communication fails, the consequences are serious:

People can’t access shelter spaces efficiently
Mental health support becomes harder to deliver
Residents feel less safe, less understood, and less empowered

What it does

Most translation tools were built for the world's dominant languages. Hearth was built for the ones left out. It's a privacy-first, on-device communication tool for shelter workers and residents - powered by Tiny Aya, a model trained specifically to serve low-resource languages like Amharic, Tigrinya, and Somali. Rather than acting as a typical translator, it serves as a low friction communication layer that supports real conversations, connection, and understanding.

Core Features

Real time communication flow from voice to text, translation, and text to speech
Automatic language detection, even when staff do not know what language is being spoken
Guided conversation prompts to help staff initiate and navigate sensitive interactions

Designed for Safety and Privacy

No reliance on third party translation APIs
No account creation or personal data storage
Built for shared device use in low resource environments
Aggressive conversation alerts to flag potential safety concerns
Saved local transcripts to help protect residents/workers from verbal abuse, with customizable retention rate

And beyond translation, we’re thinking about empowerment:

Helping residents express themselves
Helping staff ask better questions
Creating space for understanding, not just information exchange

How we built it

We made very intentional technical decisions to align with the realities of shelters.

Frontend: Next.js + TypeScript (PWA for accessibility and low friction)
Backend: Python + FastAPI
Translation Model: Tiny Aya (multilingual, supports local deployment)
Speech & Language Detection: Whisper

The key principle was removing any dependency on external services for core functionality to

Protect sensitive conversations
Ensure reliability even with limited connectivity
Support underserved languages

We also designed lightweight UI flows so that:

Zero to Minimal training is required
Causing another barrier for communication

Challenges we ran into

Balancing privacy with functionality
Orchestrating AI models for translation, language detection, text-to-voice, and voice-to-text features

Accomplishments that we're proud of

Built a fully privacy-conscious pipeline (voice → translate → voice) without relying on third-party APIs
Designed for edge cases most apps ignore (unknown languages, low literacy, no context)
Made intentional choices to ensure this could realistically be deployed in shelters

What we learned

The real challenges women in shelters face go beyond access to information and are deeply shaped by communication, trust, and daily constraints
Gained hands on experience working with a new technologies, including Aya from Cohere and Whisper
Learned how to integrate multiple AI components into a cohesive, real time system

What’s Next for Hearth

Conduct ethics study
Fine-tuning the models with trauma-informed language, shelter-specific terms, and sensitive topic handling
Safety guardrails to detect and handle harmful language responsibly
Support code-switching
Townhall mode; One-person broadcast mode
Testing directly with shelters like The Bloom Group
Explore deployment in low connectivity environments to ensure consistent performance in shelters