The problem
Ant by DataGain helps researchers and other teams transcribe and analyze recordings. My work connected the processing of long audio to the tools people use after a transcript is ready. That meant handling unusually large recordings as well as giving users a way to ask questions across their files.
What I built
I built a backend pipeline for recordings that could run 14–18 hours. It starts an EC2 worker when needed, splits audio into manageable segments, stores the pieces in S3, and sends them to Deepgram for transcription. I also developed a RAG-based chat application at both file and folder level for summaries, insights, and meeting minutes. The FastAPI service uses LangGraph and retrieval components to bring relevant transcript content into responses. The project also included pricing work and tracking of model token usage and estimated AI costs, with PostgreSQL used in parts of the backend.
- The technical challenge was preserving a useful path from a lengthy recording to an answer a person can inspect. Breaking up the audio makes processing practical; organizing and retrieving transcript content makes the resulting material useful beyond a raw text export.
