Assistary.AI
Automated Minutes of Meeting & Action-Item Extractor
An internal Astheron product for teams that spend the day in meetings.
Case Study
The problem
Minutes get written by hand by the person who should be thinking, not typing. The result arrives late, reads differently from meeting to meeting, and loses the part that decides everything: who committed to what, and by when.
Architecture
Next.js and NestJS handle the app and API, with a Python microservice for audio processing. Recordings enter a Redis queue and are processed in the background — transcription with speaker identification first, then an LLM extracts decisions and action items from the already-structured transcript.
The hard part
Meeting audio files are large and slow to process. Running that inside the request cycle means the user waits on a blank screen until the request times out — so processing had to be asynchronous from the initial design, not optimised in after the problem appeared.
Outcome
Meeting recordings become structured minutes with speaker identification, complete with action items, owners and deadlines — extracted automatically rather than retyped.
System Architecture
App & API
Next.js · NestJS
Handle the application and the API.
Queue
Redis queue
Recordings enter the queue and are processed in the background.
Python microservice
Audio processing.
AI path
Transkripsi + diarisasiAI
Transcription with speaker identification, first.
LLMAI
Extracts decisions and action items from the already-structured transcript.
Why it is shaped this way
The order is the point: the transcript is structured first, and only then does the LLM work on top of it.
Engineering Decisions
End-to-End Transcription & Summarisation
Turns meeting recordings into structured text with speaker diarization.
Actionable Insight Extraction
An optimised LLM automatically extracts agreements, action items, owners (PIC) and deadlines.
Asynchronous Job Queue
A Redis queue processes large audio files in the background without blocking user requests.