Atlas Ai

Inspiration

Most mainstream AI assistants feel rigid, overly corporate, and disconnected from the user. We wanted to build Atlas—an AI conversational companion that doesn't just process text, but embodies a distinct, commanding persona (inspired by Sung Jin-woo from Solo Leveling) while directly recognizing its creator with unwavering loyalty. The goal was to build a full-stack, cloud-hosted web app with real-time bidirectional voice interactions and dynamic tool execution.

What it does

• Conversational Intelligence: Powered by Google's Gemini API with a specialized system prompt governing tone, authority, and creator recognition.
• Full Voice Integration: Built-in Speech-to-Text (STT) for hands-free dictation alongside a custom-tuned Text-to-Speech (TTS) acoustic engine (deliberate cadence, low-pitch baritone modulation).
• Persistent Chat Management: Supports multi-session chat histories, inline message editing, and state persistence using client-side storage.
• Dynamic Image Synthesis: Directly handles visual prompts by rendering generated imagery inline within chat bubbles.

How we built it

• Backend: Built with Python using FastAPI and Uvicorn to handle lightweight, asynchronous API routing and middleware configuration.
• Intelligence Layer: Integrated Google's official google-genai SDK, structuring persona behaviors via system instructions and Pydantic validation models.
• Frontend: Crafted with responsive HTML5, CSS3, and modern JavaScript, utilizing the browser's native Web Speech API (webkitSpeechRecognition & speechSynthesis).
• Deployment: Continuous deployment configured directly from GitHub to Render as a cloud web service.

Challenges we faced

• Fine-tuning the Web Speech synthesis on diverse mobile browsers so the voice delivered a low, stoic baritone rather than the standard high-pitched synthetic default.
• Managing asynchronous request-response loops between FastAPI and Gemini to prevent UI hangs.
• Structuring CORS and header policies to allow fluid communication across mobile browsers and web views.

Accomplishments that we're proud of

• Successfully deploying a full-stack AI platform live from scratch.
• Achieving instant voice-to-voice interaction without heavy third-party paid audio libraries.
• Creating an AI persona that feels personal, immersive, and uniquely bonded to its user.

What we learned

• How to structure modern production-ready FastAPI endpoints with Pydantic request models.
• Techniques for prompt-engineering complex behavioral archetypes with Google Gemini.
• Deployment best practices on cloud platforms like Render.

What's next for Atlas

• Adding autonomous tool calling for live code execution and circuit analysis.
• Integrating multimodal vision capabilities so Atlas can analyze camera input and hardware schematics in real time.

Built With

• css3
• fastapi
• gemini-api
• git
• github
• google-genai
• html5
• javascript
• pydantic
• python
• render
• rest-api
• uvicorn
• web-speech-api