Settle

Shows moderate and high risk clauses
LLM-generated description for a specific clause and why it matters
submitting an image to analyze
Utilizing the chatbox to learn more about the clause
Showcasing the original text for the clause that was analyzed

Inspiration

We were inspired by how often people we have signed legal documents without even thinking about reading the fine print. For us, it was because legal jargon created a barrier where even if we wanted to understand the clauses, we would not be able to. And in some cases, this could easily lead to accruing hidden fees and penalties, or signing off on unfair terms. We wanted to reduce this information gap and give users clarity and confidence before they sign anything.

What it does

Running fully on-device, Settle is an AI-powered document scanner that identifies and highlights risky clauses in real time. It translates complex legal language into simple explanations. It provides a color-coded risk system (red for high risk, yellow for caution) so users can quickly understand what matters the most, and also allows a chat box for users to ask more questions about any given clause.

How we built it

The app uses camera-based text detection to extract clauses into text boxes through a bitmap. Once the textboxes have been created, we use a 2-step verification system to filter irrelevant text before using a local LLM inference to analyze risk. We combined keyword detection with contextual verification to improve accuracy and avoid over-flagging.

Challenges we ran into

One major challenge was ensuring accurate risk detection without over-relying on keywords. Many legal terms depend heavily on context, so we had to design a second layer of validation to ensure that everything being highlighted is actually useful information. We also faced constraints around running models efficiently on-device while maintaining fast response times. To address this, we shifted our approach from broadly detecting all “good,” moderate, and risky text to focusing only on truly problematic clauses, reducing noise and also improving precision to make the output far more actionable for the user.

Accomplishments that we're proud of

We were able to successfully build a fully on-device pipeline that performs real-time contract analysis without sending any data to the cloud. We also created a clear and intuitive UI that makes complex legal information easy to understand within seconds, ultimately curating an application that genuinely seems useful to society.

What we learned

We learned how to optimize AI models for on-device performance while also balancing accuracy with low latency. In the end, we were able to quickly design user-centric explanations for complex domains like legal text. We also gained experience working with real-time computer vision and NLP together.

What's next for Settle

We plan to expand support for full document uploads alongside enabling multi-document sessions so people can upload larger and more documents at the same time. We are also in the process of integrating a local vector database for smarter follow-up questions.

NOTE: APK is published as a release in our GitHub repository. https://github.com/anuragchillarige/team-bees/releases/tag/v1.0.0