case lens
@tejasgk|2025
Building an AI Evidence Analysis Tool for Lawyers
The project we built at the hackathon was essentially an AI-powered evidence analysis system for lawyers.
The idea came from a pretty straightforward problem: legal cases can generate an absurd amount of evidence.
A lawyer might receive:
- Hours of video footage
- Thousands of WhatsApp messages
- PDFs and documents
- Images
- Audio recordings
- Other case-related files
The problem isn't necessarily getting access to the evidence. It's finding the useful information buried inside it.
Imagine receiving a WhatsApp export containing 50,000 messages. Somewhere inside those messages might be a conversation that's extremely important to the case.
You know it's probably there.
You just don't know where.
The same problem exists with video. A case might contain ten or twenty hours of footage, but only a few minutes could actually matter.
Our idea was to let lawyers upload all of this evidence into one system and then interact with it almost like they were talking to someone who had already gone through everything.
Upload the Evidence
The first part was simply getting the evidence into the system.
Instead of treating everything as plain text, the system needed to understand different types of information.
A WhatsApp export could be parsed into conversations.
Documents could have their text extracted.
Videos could be transcribed and processed.
Audio could be converted into searchable transcripts.
Everything would eventually become information that the system could understand and search through.
Ask Questions About the Entire Case
This was the part I found the most interesting.
Once the evidence had been processed, the lawyer could simply ask questions.
For example:
"When did they first discuss the payment?"
Or:
"Find every time Prasanna was mentioned."
Or:
"What happened in this video?"
Or even:
"Summarize everything related to the agreement."
Instead of manually searching through thousands of messages or watching hours of footage, the system could search through the uploaded evidence and return the relevant information.
The important part was that it wasn't supposed to just generate a random AI answer.
The answer needed to come from the actual evidence that had been uploaded.
Ideally, every answer would also point back to its source.
If something came from a WhatsApp conversation, you should be able to see the messages.
If something came from a document, you should be able to open the relevant section.
If something happened in a video, you should eventually be able to jump directly to that timestamp.
That makes the AI less of an "answer machine" and more of an interface for navigating evidence.
Using Sarvam
Since Sarvam was one of the sponsors of the hackathon, we had to incorporate their API into the project.
That actually fit the problem pretty naturally because a huge part of dealing with real-world evidence is converting messy, unstructured information into something machines can work with.
Especially in India, evidence isn't necessarily going to be perfectly formatted English text.
You might have conversations involving different languages, audio recordings, videos and people switching between languages in the same conversation.
Being able to process that information before searching and reasoning over it becomes incredibly useful.
The Bigger Idea
What we managed to build during the hackathon was obviously just a prototype.
But I think the underlying idea could go much further.
Imagine opening a case and seeing:
People involved
Important dates
Timeline of events
Payments mentioned
Locations
Contradictions
Relevant conversations
Supporting evidence
And then being able to ask questions across all of it.
Instead of navigating evidence through folders like:
Evidence/WhatsApp/chat-final-2.txt
you'd navigate it through questions.
"Show me everything related to the ₹5 lakh payment."
"What happened between March 12 and March 18?"
"Where do these two people contradict each other?"
"Show me every piece of evidence supporting this claim."
The AI doesn't replace the lawyer or decide what the evidence means legally.
It just makes an enormous pile of information searchable, understandable and navigable.
That's what I liked about the project.
It wasn't really about adding AI to legal software.
It was about taking something humans are terrible at doing manually, going through ridiculous amounts of information, and giving them a much better interface for doing it.