A developer at Uralsib bank created an AI secretary robot to automate meeting summaries. The system operates entirely offline within the bank's closed network to ensure security. The workflow involves a custom recorder written in Go that connects to WebRTC-based conference rooms, captures audio, and logs participants' activity with millisecond precision.
For transcription, the developer used the OpenAI Whisper model, specifically the ggml-large-v3-turbo version, which requires 1.5 GB of memory. By utilizing a GPU (GeForce GTX 1080), the processing speed was increased to a 4:1 ratio, allowing a one-hour meeting to be processed in 15 minutes.
The final stage uses YandexGPT-5-Lite via Ollama to analyze the transcribed text and generate concise meeting summaries. The entire infrastructure is managed by a central orchestrator written in Go using a Postgres database for task queues. A pilot group of 20 people was used to test the system, processing 3 to 5 meetings per day.
