VNPT AI Hackathon 2025

MeetMate SAAR

Self-Reflective Agentic RAG Engine for Meeting Intelligence (Pre-In-Post)

VNPT AI Hackathon 2025 - 2nd Runner-up
In production rollout at LPBank
View DemoRead Proposal PDFGitHub
VNPT AI Hackathon team

Overview of MeetMate

MeetMate is an AI meeting assistant powered by ASR, RAG, LLMs and tool-calling, with VNPT APIs & Cloud services. It follows the entire meeting lifecycle to automate minutes, actions and knowledge, helping teams make decisions faster and with greater transparency.

Three core blocks

Pre-Meeting - Intelligent preparation

  • Syncs calendars from Outlook/Teams and understands topic, project and participating units.
  • Uses RAG over internal documents to pull relevant policies, proposals and past minutes.
  • Generates a suggested agenda and pre-read pack and sends it to participants in advance.

In-Meeting - AI co-host during the meeting

  • A MeetMate bot joins Teams/desktop meetings with a "Live Notes - Actions - Ask AI" panel.
  • Provides real-time transcription (multi-language) by speaker, with live recaps along the timeline.
  • Detects Actions / Decisions / Risks and suggests creating tasks, scheduling follow-ups, and attaching documents.

Post-Meeting - Turning discussion into execution

  • Produces standardized professional Minutes of Meeting: objectives, key discussion points, decisions, actions/owners/deadlines and risks.
  • Syncs tasks to tools such as Planner/Jira/Work/internal PM systems and supports reminders.
  • Stores transcripts, highlights and meeting knowledge for search and Q&A after the meeting.

Core Features Across the 3 Stages Breakthrough

Pre

Before the meeting "Walk into the room already aligned"

  • MeetMate sends a reminder with objective, attendees and project context.
  • Automatically assembles a pre-read pack from relevant policies, proposals and previous minutes.
  • Participants log questions/risks/requests in advance; the meeting owner sees a consolidated view to refine the agenda.
In

During the meeting "An AI co-host that listens, understands and suggests"

  • The MeetMate bot joins with a "Live Notes - Actions - Ask AI" panel, quietly updating the sidebar instead of interrupting the discussion.
  • Delivers real-time, speaker-aware transcription and rolling recaps, so everyone (including late joiners) instantly sees who said what and what's been agreed.
  • Detects actions, decisions & risks in what people say using intent understanding, converts them into structured ADR entries and tool suggestions (tasks, follow-ups, open documents), and answers questions powered by LightRAG that prioritises the current meeting context and internal documents - with every action gated by one-click human confirmation.
Post

After the meeting "Minutes are done and work doesn't get lost"

  • Instantly generates standardized Minutes of Meeting (summary, decisions, actions, risks with timecodes).
  • Pushes action items into Planner/Jira/Work/PM tools and assigns them to the right owners.
  • Each user gets a personal "My Recap" and can later query: deadlines, decisions, or where something was approved across past meetings.

Full-stack ArchitectureEnd-to-end workflow & system mapOrchestration, data flow, and stage routingDeployment layers & enterprise integrations

MeetMate full-stack architecture slide 1
MeetMate full-stack architecture slide 2
MeetMate full-stack architecture slide 3

MeetMate | SAAR - Stage-aware Agentic Adaptive RAGAI Architecture Focus

MeetMate SAAR architecture overview diagram

Pre-meeting RAG & agenda prep

History-aware RAG builds agenda and pre-read packs before the meeting.

  • Ingest calendar context, series history, pending actions/risks, and meeting attachments.
  • Retrieve project/topic docs + policies with priority scoring (recency + unresolved ADR boosts).
  • Generate a 3-7 item agenda with time allocation and cited pre-read links; collect attendee questions/requests.
Pre-meeting AI architecture diagram

In-meeting realtime multi-agent RAG

Low-latency LangGraph pipeline with VNPT APIs turns speech into recap, ADR, Q&A, and governed tool-calls.

  • ASR streaming + stage/event router; tick scheduler controls LLM budget.
  • Shared state: transcript window, topics, ADR, Q&A, tool suggestions, and metrics.
  • Three-layer retrieval: meeting context → project/topic docs → global knowledge, always with citations.
In-meeting AI architecture diagram

Post-meeting long-context consolidation

Refines ADR and produces minutes with compliance-ready archiving.

  • Consolidate transcript/recap/ADR, dedupe decisions, and fill owners/deadlines with timecodes.
  • Enrich with policies and historical minutes using weighted long-context RAG.
  • Generate MoM + highlights, sync tasks to Planner/Jira, and archive with audit trails.
Post-meeting AI architecture diagram

Deployment & Stack

Production stack

  • Backend: Python services + REST/WebSocket API.
  • Data: PostgreSQL OLTP + vector store for RAG.
  • Infra: VNPT Cloud deployment with audit logging.
  • LLM providers: multi-model routing by SLA.

Results & Impact

2nd Runner-up at VNPT AI Hackathon 2025
Top 3 among 200+ teams nationwide.
Production rollout at LPBank
Private SaaS backend with audit-ready logs.
Evidence-first responses
Citations required; no-source -> no-answer policy.

Contact

Want to see the full system design or deployment details? Reach out and I will share the deep-dive docs.

Send emailGo to main contact
Ask me