GetLevrg Meeting Intelligence
Meeting transcripts arrive from Google Drive, a two-stage LLM pipeline pulls out the action items, risks and milestones and scores the call against a fixed rubric, and every meeting gets a chat tab that answers questions from its own transcript.
The challenge
What we set out to solve
Client meetings ended as DOCX exports in a Drive folder. Nothing was scored, nothing was searchable, and an account manager who wanted to know whether a relationship was improving or sliding had to read transcripts end to end to find out.
Our approach
How we thought about it
Two stages, deliberately split. The first parses title, attendees and transcript with plain regex, so the cheap work costs nothing; the second runs extraction and then rubric scoring through an LLM. Transcripts are chunked into 150–600 token speaker-aware passages and embedded into a Chroma collection of their own, per meeting, so an answer is grounded in that call and cannot drift into another one. The LLM and the embedding model are both chosen by an environment variable, whether Groq, Gemini or a local Ollama, which is what keeps the cost model open.
What we built
The solution
Meetings ingest on a timer, score themselves out of ten, and land in a review queue that belongs to each user rather than to the team. Analytics track the score over time by client and by CSM, the chat tab answers from the transcript, and a command palette jumps to any meeting, client or page. The whole stack of API, app, Postgres and Chroma comes up with one docker compose command.
Illustration of Northwind Agents, a system of this kind built by ProbahiCraft, using sample data.
