Homi | Devpost

Inspiration

A big part of our inspiration was taken from AirBnb and the features and issues that the app had solved. Yet, when re-imagining it as if it were built in 2025 within the big age of AI, it felt intuitive to have incorporated AI into streamlining the grueling process of real estate. From a consumer side, it's often difficult to parse through thousands of different listings just to be able to find someplace safe, affordable, and available for you to stay at during a vacation. From a seller's side, the process ranging from having to manually obtain sample photos of your property to deciding how much to charge when you have zero prior experience in selling a property, is both challenging and difficult.

What it does

Homi makes it easy for guests to find safe, affordable stays and helps hosts showcase their properties to maximize profit through the following features...

AI agents curating personalized, potential homes for consumers to stay at
AI agents automating the listing process, removing any headache from a first time seller.

How we built it

Homi is built entirely on a FastAPI backend and Next.js frontend. The backend itself runs on uvicorn with REST API + Websocket endpoints.

Core Technical Implementations

For usage of real-time object detection we employ the use of YOLOv8n. An open-source model coming from a large family of similar object detection usage. This model runs server-side with session-based aggregation storing detections in-memory with frame metadata. We also use this to capture and later transmit images as data URIs over WSS.

We employ Groq llama 3.3 70b to extract structured search params (location, dates, guests, price) from natural language. A state machine then tracks missing params, and triggers follow-up questions through the use of Anthropic's Claude Sonnet 4.5. Once complete, params trigger Elasticsearch hybrid search with relevance threshold filtering.

Another key feature is our Tinder-style swiping interface. Using Framer Motion gesture handlers on React client. each swipe tracked to Supabase + Letta for preference learning. Saved listings auto-ranked using Claude with contextual reasoning (price drops, preference alignment, availability, etc.)

Challenges we ran into

AirBnb itself doesn't offer open API keys to students and therefore, we had to be clever in the ways in which we obtained the listing data.

We had to pivot a couple of times. Beginning from an AR-powered application, we were unfortunate in our attempts to try and automate machine learning workflows on-top of our hardware and obtain outputs in a server-less environment. We then centered ourselves around an AI-native approach through the use of open-source models and AI agents.

Accomplishments that we're proud of

Creating a fully, custom designed, suite. One that incorporates elements from every single one of us meaning that this project is entirely our own.

Arriving to a finished product, as a deliverable isn't always promised despite that being the entire point of Hackathons. Doing this with two first-time Hackathon attenders nonetheless!

What we learned

Lots, and lots about hardware. Hardware, in the form of AR glasses, served as one of the main axioms and previous "invariants" to our product as it enabled custom pricing for listings as well as customer previews; having a hand in both buyer + seller workflows.

Also, since one of our teammates had a background as a designer we really could delegate a lot more time into making a User-focused design + maintain a Product-focused mindset. These two paired together resulted in a beautiful UI made with users in mind to appease their life and make using our application all that much easier.

What's next for Homi

Globalization! Currently Homi is only available for US locations. A future plan that we've all thought about is how we can globalize Homi for any person hoping to find a nice place to stay without all of that headache associated.