Hello response
Default screen
Complex question response

IKARIS - HackUTD 2025

Inspiration

CRE analysts are surrounded by a mess of PDFs, spreadsheets, and an endless amount of tabs. We wanted to help solve this by creating an AI model called IKARIS to process possible leasing risk, OPEX spikes, and provide trustworthy answers quickly.

What it does

Answers complex questions by pulling text data across contracts, reports, and CSVs
Create efficient predictive analysis based on data trends
Provide confidence and inline sources for why it chose the answer which provided ## How we built it
Our main agent is powered by NVIDIA's Nemotron, which we trained with different machine learning algorithms in order to provide CBRE context for analysis
For back-end, we utilized Python with flask, as well as multiple ML libraries such as scikit learn, LangChain, and chroma.
For front-end, we used Node.js and React ## Challenges we ran into
We needed to find the best method to train our model on unstructured data (PDFs)
We were struggling to find property information data sets or free APIs that we could train our model with.
Since our original model was only introduced to the above data, it only responded with memory of that respective information. We had to introduce some conversational ability for this reas. ## Accomplishments that we're proud of
For the unstructured data issue, we used RAG to chunk data to provide enough context for model without overfitting the model with unnecessary information
Ultimately, for the structured data issue, we assessed our needs for data and created an algorithm which would pull directly from CBRE quarterly data to create similarly structured and trending synthetic sample data.
We're really happy with the fact that the Chatbot is hybridized based off the two main ways that we're training it, as well as being able to be conversational using basic chat training. ## What we learned
RAG and small models can beat a single large model when the data is niche and is focused on a specific context
Crafting a clear and specific prompt gives the chatbot clear goals and constraints, keeping it on topic, aligned with the primary use case. ## What's next for IKARIS
We have room for contextual scalability and accuracy in terms of the data being fed into the model. Currently we have have few CBRE pdfs of their reports and a generated dataset built around what the company might use or need. If/when CBRE were to actually use this, they can provide and utilize any internal reports and data they may need. That way, this chatbot can be the perfect assistant for CBRE professionals.
Additionally, the front end can be built into any existing software or expanded to include a more dynamic interface.