# Lazarus

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Vishwanath Archak M R](https://cerebralvalley.ai/u/vishwanatharchak)
- **GitHub:** https://github.com/VishwanathArchakMR/Sahaay
- **Demo video:** https://youtu.be/Bu_sdsKfEvA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3705

I built Sahaay, a voice-first AI assistant for Windows designed to make everyday computer use easier for older adults and people who find traditional digital interfaces difficult to navigate.

The idea behind Sahaay is simple: users should be able to explain what they want in natural language instead of learning complicated software or navigating multiple menus. Sahaay understands the request, plans the required steps, and helps carry them out across local files, browser tabs, websites, and Windows applications.

Sahaay has two primary modes: Automate and Help. Automate allows users to perform computer tasks through natural instructions, while Help provides a conversational interface for asking questions and getting assistance.

For complex tasks, I integrated Meta Muse Spark 1.3 as the primary reasoning and planning model. It analyses user objectives, breaks them into smaller steps, identifies the required tools, and coordinates task execution. Muse Voice Transcribe 1.0 supports voice-based interaction, with language hints for Kannada, Hindi, and English. I also integrated Muse Wonder 1.0 and SAM 3.1 for visual reasoning and screen-element grounding.

The system combines these models with a practical computer-use architecture:
- Electron, React and TypeScript for the Windows desktop application.
- Python and FastAPI for the local backend and orchestration.
- Chrome/Edge Manifest V3 extension with a Native Messaging bridge for browser interaction.
- .NET Windows companion using Windows UI Automation for native desktop controls.
- SQLite for local application data and task records.
An important part of Sahaay is that it does not simply trust the model's proposed actions. Its execution workflow follows Observe → Propose → Authorize → Execute → Verify → Record. An Action Gate controls whether an action is permitted, and an independent verification layer checks the outcome before the system reports completion.

I demonstrated Sahaay through real workflows, including organising downloaded files using a Kannada instruction, researching an iPad on Amazon within a specified budget, finding and playing a YouTube performance, analysing a webpage and creating a report in Notepad, and reading and executing a reply to an internship email.

My goal with Sahaay is to move beyond chat-based AI and build an assistant that can understand people's intentions and help them complete meaningful tasks on their computers, while keeping user control and authorization central to the process.

## More from Meta Global AI Developer Hackathon

- [Witness](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3702)
- [Aftermath](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3703)
- [MoeCode](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3704)
- [PhysLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3706)
- [Jayseattle](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3707)
- [QrMap](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3708)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3705. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt