Atlas-Talks

Inspiration

We’ve all used map apps that work, but don’t really listen. You ask for help, and they just follow commands. They don’t understand the context, the conversation, the emotion, or what you actually mean. That’s where the idea for Atlas-Talks was born.

We wanted to create something that feels alive! And we wanted to go after something that doesn't exist in almost any form in the industry.

What it does

• Atlas-Talks is an intelligent, real-time navigation system.
• It processes live data, understands conversational cues, and coordinates responses that make sense both in chat and on the map. It can follow your requests. You can literally ask it in human language to show around the maps the locations you desire, with tons of customizations, real life cues.
• No, IT'S NOT CHATGPT. It does not rely on it. We built this ground up. Because an LLM alone cannot control a map.

How we built it

We built the entire system from scratch, focusing on speed, realism, and intelligence.

Here’s the breakdown:

• Data Synthesis & Parsing: Built our own pipeline to continuously generate and process live navigation data. Extracted insights from the data on what needs to be incorporated and what not.
• Real-Time Analysis: Every model runs locally — on CPU only — analyzing and updating routes in near real time.
• Conversational Layer: Connects what’s happening on the map to natural, human-like dialogue.
• Except for the voice assistant, everything runs in-house.
• No cloud dependencies. No shortcuts. Just raw engineering, lots of data analysis, data compilation, cleaning and a lot of debugging.

Challenges we ran into

• Making real-time processing actually real-time without GPUs.
• Syncing conversation flow with live navigation data insights. This is EXTREMELY DIFFICULT.
• Keeping latency super low while still making the system feel intelligent.
• And of course… the classic hackathon battle: sleep deprivation + endless “why isn’t this working?” moments.

Accomplishments that we're proud of

• Built a full navigation + conversation pipeline completely from scratch relying on tons of data that we generated, compiled, cleaned and parsed.
• Pulled off intelligent back-and-forth between map state and chat interface.
• Made something that doesn’t just work. it feels alive.

What we learned

• Innovation isn’t about making the flashiest AI. It’s about making it usable and practical.
• Optimization matters. We learned how much you can squeeze out of a CPU with smart pipeline design.

What's next for Atlas-Talks

• We believe this is a real world product that so far DOESN'T have MOST of the features we baked in.
• Since we meticulously went for cases that truly can be innovative, we believe we can make this into a full fledged product that actually becomes the standard for these kinds of tasks.
• Adding additional intelligence and context will be the next immediate steps!

Built With

• css
• html
• huggingface
• javascript
• ollama
• python
• transformers