AI & Machine Learning / CASE STUDY
MeetingSense AI.
AI Meeting Assistant
01 — THE CHALLENGE
The problem.
Important decisions, tasks, and discussions often get buried inside long meeting recordings. Finding one specific detail later means searching through recordings or asking teammates. This becomes even harder when meetings contain mixed languages such as English and Hinglish.
02 — THE APPROACH
A thoughtful solution.
Built an AI meeting assistant that turns YouTube recordings, audio files, and transcripts into searchable meeting knowledge. The system transcribes the meeting, breaks the content into smaller sections, stores it for fast search, and lets users ask questions about what was discussed. It can also find action items, create meeting minutes, identify open questions and disagreements, prepare follow-up emails, and trace how a topic changed across multiple meetings.
03 — BEHIND THE BUILD
How it came together.
The main problem was not just converting audio into text, but making that information useful after the meeting was over. A single meeting can contain decisions, action items, questions, disagreements, and important details spread across an hour-long conversation. The system first takes a YouTube link, audio file, or transcript and converts it into clean text. The text is then split into smaller sections and stored in Pinecone so the relevant part can be found quickly later. A LangGraph-based agent chooses the right tool depending on the user's question. For example, it can search the transcript for a specific discussion, extract action items, generate meeting minutes, find unanswered questions, or search across older meetings to understand how a decision changed over time. English and Hinglish meetings are handled through different speech-to-text services. The result is a practical workspace where past meetings become easy to search and understand instead of becoming forgotten recordings.
04 — THE RESULT
Searchable AI Meeting Assistant
Turned long meeting recordings into searchable information, allowing users to quickly find decisions, tasks, discussions, and unresolved questions without replaying the entire meeting.
Screens, systems,
and the details.
06 — DEVELOPMENT MILESTONES
The build, step by step.
- 01
Built the FastAPI backend and React frontend with authentication and meeting management
- 02
Added support for YouTube links, audio files, and text transcripts
- 03
Built the transcription pipeline using Deepgram for English and Sarvam AI for Hinglish
- 04
Added text cleaning, chunking, embeddings, Pinecone vector search, and meeting-level retrieval
- 05
Built the LangGraph agent with tools for Q&A, summaries, action items, minutes, follow-up emails, open questions, and disagreements
- 06
Added cross-meeting search, caching, audit logs, Supabase Row Level Security, Docker support, and deployment configuration
07 — LESSONS & CONSTRAINTS
What made it interesting.
Handling English and Hinglish meetings with different speech-to-text systems.
Searching large meeting transcripts without sending the entire transcript to the AI model.
Preventing the AI from making up owners, deadlines, or decisions that were never mentioned.
Keeping user and meeting data separated when many users use the same backend service.
Managing slow or unreliable external services such as transcription, embeddings, and LLM APIs.
Keeping repeated queries fast with server-side and client-side caching.