Skip to content
All workCase study 02 / 04

P—022025Live demo

Meet.AI

Video meetings where an AI agent joins the call, then writes the notes.

Role
Solo developer — design, full stack, deployment
Type
Full-stack SaaS · Personal project
Focus
Event-driven AI processing
Year
2025
Meet.AI landing page with the headline ‘AI Meeting Assistant for Transcription, Summaries & Smart Notes’
fig. P02 — Meet.AI landing page with the headline ‘AI Meeting Assistant for Transcription, Summaries & Smart Notes’

§ 01Overview

Meet.AI lets a user define AI agents with their own instructions, schedule meetings, and run video calls in which the agent takes part as a participant.

When the call ends, the recording and transcript are processed in the background into a markdown summary with timestamped notes. Afterwards the user can keep chatting with the agent, which answers with the meeting summary as context.

§ 02Problem & solution

Problem

Meeting notes are tedious and inconsistent, and context disappears once the call ends. Building a fix means coordinating real-time video, slow and bursty AI work, and paid tiers — without making the app feel slow.

Solution

Treat each call as a stream of events. Stream webhooks drive the meeting lifecycle, heavy AI work runs as retryable Inngest steps outside the request path, and every result lands back in Postgres where typed tRPC procedures serve it to the UI.

§ 03My role

Solo developer — design, full stack, deployment. What I personally implemented:

  1. 01Modelled users, agents and meetings in a Drizzle schema on Neon Postgres, with typed tRPC routers per module (agents, meetings, premium).
  2. 02Wired Stream webhooks into the meeting lifecycle: session start connects the agent through OpenAI Realtime; session end, transcription-ready and recording-ready events update the meeting.
  3. 03Built the Inngest pipeline that fetches the transcript, resolves each speaker to a user or an agent, and summarizes it with Gemini 2.5 Flash through Inngest Agent Kit.
  4. 04Implemented post-meeting chat: new messages trigger a background job that answers using the agent's instructions plus the stored summary.
  5. 05Set up Better Auth (email/password, Google, GitHub) and Polar subscriptions with free-tier limits of three agents and three meetings.
  6. 06Deployed on Vercel.

§ 04Architecture

fig. A02 — Architecture
  1. L1Client
    • Next.js 16 App Router
    • TanStack Query
    • Stream Video & Chat SDK
    typed queries & mutations
  2. L2API
    • tRPC routers
    • Better Auth routes
    • Webhook handler
    events from webhooks & procedures
  3. L3Async
    • Inngest: meetings/processing
    • Inngest: chat/message.new
    model calls per step
  4. L4AI
    • Gemini 2.5 Flash · summaries
    • OpenAI Realtime · in-call agent
    results persisted
  5. L5Data & services
    • Neon Postgres + Drizzle
    • Polar billing
    • Stream
  • Nothing slow happens inside a request: webhooks only record state and emit events.
  • Each Inngest step (fetch → parse → add speakers → summarize → save) retries independently.

§ 05Technology stack

Frontend
Next.js 16ReactTypeScriptTailwind CSS 4Radix UInuqs
API & auth
tRPCTanStack QueryZodBetter Auth
Data
Neon (serverless Postgres)Drizzle ORM
AI & jobs
Gemini 2.5 FlashOpenAI Realtime (via Stream)Inngest + Agent Kit
Platform
Stream Video & ChatPolarVercel

§ 06Key features

  • F01

    Custom agents

    Create agents with their own instructions and generated avatars.

  • F02

    AI in the call

    The agent joins the video call as a participant through OpenAI Realtime.

  • F03

    Automatic summaries

    Transcripts become markdown overviews with timestamped notes.

  • F04

    Recording & transcript

    Each meeting keeps its recording and a searchable transcript.

  • F05

    Ask the meeting

    Chat with the agent afterwards; it answers from the meeting summary.

  • F06

    Plans & limits

    Free tier capped at three agents and three meetings; upgrades via Polar.

§ 07Engineering challenges

C1

AI work that outlives a request

Problem

Fetching, parsing and summarizing a transcript can take far longer than a serverless request is allowed to run.

Approach

Webhooks only persist state and emit events. Inngest runs the pipeline as separate steps, so a failed model call retries without refetching the transcript.

C2

Who said what

Problem

Transcript lines carry raw speaker IDs that can belong to either a human user or an AI agent.

Approach

Collect unique speaker IDs, query both the users and agents tables, and merge them before the transcript reaches the summarizer.

C3

A call is a state machine

Problem

Session start, participant leave, session end, transcript ready and recording ready arrive as separate, asynchronous webhooks.

Approach

Map each event type to one meeting-status transition in a single handler, keeping the lifecycle explicit and easy to debug.

§ 08Screens

Meet.AI landing page
S1 — Public landing page (live at meetai-assistant.vercel.app)
Next case study / P—03Book Review PlatformA Spring Boot API with JWT auth, roles and a React client.