Describe your project

1. In-Scope: For the prototype, we have demonstrated the capability of NyAI for augmenting legal chatbots with GenAI. This includes - querying courtroom judgements (using Google's NotebookLM powered by Gemini 1.5 Pro), querying courtroom proceedings (using LiveLaw reportage with Gemini 1.5 Flash), Translation into Indian Languages (including languages such as Khasi and Tulu, which are not covered under the Eighth Schedule), Text to Speech for Visually Impaired users and generating emails/petitions/summons while restricting queries not related to law.
   Pitchdeck: https://tinyurl.com/32h3bpny
   One Minute Explainer -
2. Out of Scope: Translation of handwritten land records/petitions, integrating with DPI (endpoints are not available), validating the truth with human in the loop and automatically fetching related judgements (they need to be manually uploaded). Link to NotebookLM is not shared publicly, and will not work on the test link below.
3. Future Opportunities: Integrating with Bhashini & fine-tuning responses with Gemini 1.5, translating images, crawling existing judgements to create a graph of related judgements and citing information (to serve as legal precedents), measuring the accuracy of responses, directly sourcing judgements from eSCR & statistics from National Judicial Data Grid, integrating explainability & a layer of AI governance.

Challenges we ran into

1. With the limited token size in Gemini 1.5 Flash API for testing, we were unable to upload large judgements for querying, using GenAI. Furthermore, some citations were yielding misleading results or hallucinating. To mitigate this challenge, we used NotebookLM (a project by Google) powered by Gemini 1.5 Pro, that was able to query large documents, without hallucinations. Furthermore, it was able to accurately cite the references in the answers automatically, without explicit instructions, improving the quality of research.
2. Furthermore, based on our conversations with lawyers, we discovered the need for translating results for litigants who are unfamiliar with English. Due to the bulk of the courtroom proceedings being in English, Bhashini seemed like a natural choice, but it does not support languages outside the Eighth Schedule of the Constitution (such as Tulu, Khasi & Bhojpuri). As an alternative, we used Google 1.5 Flash API to translate the results across Indian languages including Khasi & Tulu, with good accuracy.
3. As Gemini 1.5 Flash is not trained to identify & block profainity/explicit content for outputs in Indian language, the results were being blocked by Gemini 1.5 Flash API. To handle this challenge, we disabled the risk flags while integrating the API, to allow the model to answer the queries in Indian langauges.