AI Product Manager for Recurring Meetings

An end-to-end system for turning recurring product meetings into structured project memory, accountable action items, and context-aware Minutes of Meeting (MoM).

This project is not just an AI note taker. The goal is to build an AI Product Manager that remembers what happened in previous meetings, tracks open questions and deadlines, and helps teams review delivery continuity over time.

Start Here

If you want the fastest path through the project, use this order:

1. README.md
2. docs/PROJECT_STATUS.md
3. docs/REVIEW_GUIDE.md
4. docs/SUBMISSION_GUIDE.md
5. benchmark/scenarios/onboarding_growth_initiative/transcripts/README.md

If you only need the essentials:

• What the project is and how to run it README.md
• What is working and what is still in progress docs/PROJECT_STATUS.md
• Where the code and dataset links are docs/SUBMISSION_GUIDE.md
• How to use the benchmark dataset and read the score benchmark/scenarios/onboarding_growth_initiative/transcripts/README.md

All-In-One Guide

If you want to share just one link, this README is intended to be enough on its own.

1. Code Access

• Repository: https://github.com/KumarSashank/AI-Product-Manager
• ZIP download: https://github.com/KumarSashank/AI-Product-Manager/archive/refs/heads/main.zip

2. Minimum Setup

Prerequisites:

• Node.js 20+
• pnpm 8+
• Docker Desktop or Docker Engine with Compose
• OPENAI_API_KEY

Minimal Docker env:

OPENAI_API_KEY=sk-your-openai-key

Run the application:

cp .env.docker.example .env.docker
docker compose --env-file .env.docker up --build -d

Open:

• Web app: http://localhost:3001
• API health: http://localhost:3002/api/v1/health

3. Recommended Product Flow

Use this path for the most reliable walkthrough:

1. Start the app with Docker.
2. Create a project.
3. Upload a transcript.
4. Generate the Minutes of Meeting.
5. Review extracted items and accountability.
6. Upload the next transcript into the same project to show continuity.

4. Dataset Access

Direct dataset links:

• Scenario folder: https://github.com/KumarSashank/AI-Product-Manager/tree/main/benchmark/scenarios/onboarding_growth_initiative
• Transcripts folder: https://github.com/KumarSashank/AI-Product-Manager/tree/main/benchmark/scenarios/onboarding_growth_initiative/transcripts
• Scenario JSON: https://github.com/KumarSashank/AI-Product-Manager/blob/main/benchmark/scenarios/onboarding_growth_initiative/scenario.json

The benchmark transcript files are:

1. 001_week1_kickoff.txt
2. 002_week2_status.txt
3. 003_week3_scope_risk.txt
4. 004_week4_replan.txt
5. 005_week5_launch_readiness.txt

5. How To Use The Dataset

Benchmark typecheck:

pnpm benchmark:typecheck

Run our stateful method:

pnpm benchmark:longitudinal -- benchmark/scenarios/onboarding_growth_initiative/scenario.json

Run the normal baseline:

pnpm benchmark:longitudinal -- --system transcript_only benchmark/scenarios/onboarding_growth_initiative/scenario.json

Compare both methods:

pnpm benchmark:compare

6. Where To Find The Score

Benchmark reports are written to:

• Host runs: benchmark/reports/
• Docker backend container runs: /app/benchmark/reports/

7. How To Read The Result

The benchmark compares:

• current_system Our method with project memory and accountability carry-forward
• transcript_only The normal baseline without prior project memory

Expected result for the built-in scenario:

• current_system: 38 passed / 0 failed
• transcript_only: 32 passed / 6 failed

Interpretation:

• higher passed is better
• lower failed is better
• if transcript_only matches or beats current_system, that means the longitudinal reasoning has regressed

8. Current Product Maturity

Most reliable:

• transcript upload
• contextual MoM generation
• project-level action items
• benchmark evaluation

Still in progress:

• bot-based meeting joining reliability
• Chrome extension multi-speaker transcript attribution
• audio transcription experimentation

What This Project Does

The system supports three capture paths:

• Transcript upload The most reliable path today. Upload an existing transcript and generate contextual MoM, extracted items, and longitudinal accountability.
• Join with bot A Playwright-based Google Meet bot that can join meetings and capture live captions. Useful for demos and experimentation, but still reliability-limited by meeting permissions, waiting rooms, and Google auth flows.
• Chrome extension A browser-based Meet capture path that is under active development. Audio capture is supported. Speaker-attributed multi-person transcript extraction is still being improved.

Why This Matters

Typical AI meeting assistants summarize a single meeting well, but they often lose continuity across weeks. This system is designed to answer harder project questions:

• What was promised last week, and is it still open?
• Which owner missed a deadline without giving a status update?
• Which product question is still unresolved?
• Did the team actually close the launch blocker they discussed earlier?
• Is the project becoming more ready, or just generating more meeting notes?

Core Outcomes

• Persistent meeting memory across a project
• Context-aware MoM generation
• Structured extraction of action items, decisions, blockers, and open questions
• Accountability tracking by owner, team, priority, and due date
• A benchmark harness that compares the stateful system against a transcript-only baseline

Current Product Status

What is strongest today:

• Transcript upload workflow
• Context-aware MoM generation
• Project-level item tracking and accountability
• Benchmark and research evaluation scaffolding
• Dockerized local setup

What is still in progress:

• Bot-based meeting joining reliability
• Chrome extension multi-speaker speaker attribution
• Audio-transcription experimentation beyond transcript/caption capture

Usage note:

• For the most reliable flow, use transcript upload.
• Treat bot join and extension capture as experimental/preview features.

See docs/PROJECT_STATUS.md for a fuller summary of what has been achieved and what is still being improved.

Monorepo Structure

AI-Product-Manager/
├── packages/
│   ├── ai-backend/        Fastify API, AI pipelines, database access
│   ├── web/               Next.js dashboard and project workspace
│   ├── bot-runner/        Playwright-based Google Meet bot
│   ├── chrome-extension/  Browser-based Meet capture
│   └── shared/            Shared schemas, contracts, and types
├── benchmark/             Longitudinal benchmark harness
├── docs/                  Project, research, and operational docs
├── docker/                Dockerfiles and runtime setup
└── scripts/               Utility scripts

Tech Stack

• Next.js and React for the dashboard
• Fastify and TypeScript for the backend
• PostgreSQL and Drizzle for persistence
• OpenAI for MoM generation and structured extraction
• Playwright for the bot-based Meet capture path
• Chrome Extension APIs for in-browser Meet capture

Quick Start

Prerequisites

• Node.js 20+
• pnpm 8+
• Docker Desktop or Docker Engine with Compose support
• OPENAI_API_KEY

Recommended Setup: Docker

This is the easiest path for local setup and evaluation.

cp .env.docker.example .env.docker
docker compose --env-file .env.docker up --build -d

Then open:

• Web app: http://localhost:3001
• API: http://localhost:3002
• Health check: http://localhost:3002/api/v1/health

Useful commands:

pnpm docker:up
pnpm docker:down
docker compose --env-file .env.docker logs -f
docker compose --env-file .env.docker up -d --build ai-backend
docker compose --env-file .env.docker up -d --build web

Full guide:

• docs/DOCKER_RUN.md

Local Development

pnpm install
pnpm dev

Common package-level commands:

pnpm --filter @meeting-ai/ai-backend dev
pnpm --filter @meeting-ai/web dev
pnpm --filter @meeting-ai/bot-runner dev

Local environment details:

• docs/ENVIRONMENT.md
• docs/README.md

Environment Variables

Required For The Normal Demo Flow

Variable	Required	Purpose
OPENAI_API_KEY	Yes	Generates contextual MoM and structured extraction

Included With Safe Local Docker Defaults

Variable	Required	Purpose
JWT_SECRET	No for local Docker	JWT signing secret
COOKIE_SECRET	No for local Docker	Cookie signing secret

Optional, Depending On The Capture Method

Variable	Required	Purpose
GOOGLE_EMAIL	Optional	Used by the bot-runner for Google auth
GOOGLE_PASSWORD	Optional	Used by the bot-runner for Google auth
DEEPGRAM_API_KEY	Optional	Only needed for audio transcription experimentation

Minimal .env.docker

OPENAI_API_KEY=sk-your-openai-key

You can leave the dev-only JWT_SECRET and COOKIE_SECRET defaults as-is for local Docker runs.

Code And Dataset Access

Code Access

Repository:

• GitHub: https://github.com/KumarSashank/AI-Product-Manager

Download options:

• Clone:

git clone https://github.com/KumarSashank/AI-Product-Manager.git

• Download ZIP: https://github.com/KumarSashank/AI-Product-Manager/archive/refs/heads/main.zip

Dataset Access

The evaluation dataset used by this project is included in the repository under benchmark/scenarios/.

Primary benchmark dataset:

• Scenario definition: benchmark/scenarios/onboarding_growth_initiative/scenario.json
• Transcript files: benchmark/scenarios/onboarding_growth_initiative/transcripts/

GitHub paths:

• Scenario folder: https://github.com/KumarSashank/AI-Product-Manager/tree/main/benchmark/scenarios/onboarding_growth_initiative
• Transcripts folder: https://github.com/KumarSashank/AI-Product-Manager/tree/main/benchmark/scenarios/onboarding_growth_initiative/transcripts
• Scenario JSON: https://github.com/KumarSashank/AI-Product-Manager/blob/main/benchmark/scenarios/onboarding_growth_initiative/scenario.json

See docs/SUBMISSION_GUIDE.md for a submission-ready mapping of code, dataset, and benchmark assets, and benchmark/scenarios/onboarding_growth_initiative/transcripts/README.md for a short dataset walkthrough.

How To Use The App

Recommended walkthrough:

1. Start the app with Docker.
2. Sign in to the web app.
3. Create a project.
4. Upload a transcript through the project workspace.
5. Review the generated MoM, extracted items, and project item board.
6. Open the next meeting in the same project to show continuity across meetings.

Stable demo path:

• Projects -> Upload transcript -> Generate MoM -> Review action items

Preview/experimental paths:

• Join with bot
• Chrome extension capture

See docs/REVIEW_GUIDE.md for a guided walkthrough.

Testing

Core Commands

pnpm test
pnpm typecheck
pnpm lint
pnpm format:check
pnpm pre-push

Focused Commands

pnpm --filter @meeting-ai/ai-backend test
pnpm --filter @meeting-ai/ai-backend typecheck
pnpm --filter @meeting-ai/web exec tsc --noEmit
pnpm benchmark:typecheck

Expected outcome:

• pnpm test should pass
• pnpm typecheck should pass
• pnpm format:check should pass
• pnpm pre-push should pass
• pnpm lint may still show warnings in some areas, but should not fail

More detail:

• docs/TESTING.md
• docs/TOOLING.md

Benchmark Evaluation

This repo includes a longitudinal benchmark harness that compares:

• current_system The full stateful AI Product Manager workflow with project memory
• transcript_only A baseline that only uses the current meeting transcript

Benchmark Commands

pnpm benchmark:typecheck
pnpm benchmark:longitudinal
pnpm benchmark:compare
pnpm benchmark:longitudinal -- benchmark/scenarios/onboarding_growth_initiative/scenario.json
pnpm benchmark:longitudinal -- --system transcript_only benchmark/scenarios/onboarding_growth_initiative/scenario.json

Reports

• Host output: benchmark/reports
• Docker backend container output: /app/benchmark/reports

Expected result for the built-in comparison scenario:

• current_system: 38 passed / 0 failed
• transcript_only: 32 passed / 6 failed

This benchmark is important because it demonstrates that persistent project memory outperforms raw transcript-only summarization on continuity, accountability, and final project-state checks.

More detail:

• benchmark/README.md
• docs/BENCHMARK_SLIDE_SUMMARY.md
• docs/EVAL_RUBRIC.md

What We Achieved

• Built a multi-package product around recurring-meeting intelligence
• Added contextual MoM generation grounded in prior project state
• Added action-item continuity, accountability, and due-date awareness
• Added a Notion-like item workspace for reviewing and updating work
• Dockerized the stack for easier local setup
• Added a benchmark harness for longitudinal evaluation
• Framed the system as a researchable AI Product Manager problem, not just a meeting-summary tool

What Still Needs Work

• Improve bot reliability under Google auth and waiting-room edge cases
• Improve extension-based multi-speaker caption separation
• Improve audio-to-transcript experimentation beyond raw capture
• Expand the benchmark dataset beyond the initial scenario pack
• Add stronger evidence trails and inspection tooling for why the system marked an item as open, resolved, blocked, or overdue

Documentation Map

Start here:

• docs/README.md

Most useful docs:

Document	Purpose
docs/PROJECT_STATUS.md	Achievements, limitations, and roadmap
docs/REVIEW_GUIDE.md	Guided walkthrough of the product
docs/SUBMISSION_GUIDE.md	Code and dataset access for submission
benchmark/scenarios/onboarding_growth_initiative/transcripts/README.md	How to use the benchmark dataset
docs/DOCKER_RUN.md	How to run the full app locally
docs/ENVIRONMENT.md	Environment variable setup
docs/ARCHITECTURE.md	System design and data flow
docs/TESTING.md	Test strategy and commands
benchmark/README.md	Longitudinal benchmark harness
docs/AI_PRODUCT_MANAGER_RESEARCH_PLAN.md	Product and research thesis
docs/PAPER_OUTLINE.md	Research paper outline

Packages

Package	Description
packages/ai-backend/README.md	Backend API and AI pipeline
packages/web/README.md	Web dashboard and project workspace
packages/bot-runner/README.md	Bot capture path
packages/chrome-extension/README.md	Browser capture path
packages/shared/README.md	Shared contracts and schemas

Final Note

If you are reviewing this project academically, the strongest way to evaluate it is not only by “does it summarize one meeting well?” but by “does it preserve continuity across a series of meetings and improve project accountability over time?”