Meeting Is Data — Starter Kit

Resources from the conference talk "Meeting Is Data: How to Extract Shared Context from Everyday Conversations" by Jakub Sikora, VP Engineering at Circit.

What's Here

File	Description
prompt-template.md	LLM system prompt for topic classification with structured JSON output, taxonomy injection, and similarity gating
graph-api-queries.md	Microsoft Graph API endpoints for meetings, transcripts, Copilot AI insights, attendance reports, and sensitivity labels
pipeline-skeleton.md	Data pipeline pattern (ADF/Airflow) — 18-activity flow with error handling and cost optimization
privacy-checklist.md	5-layer privacy framework including EU AI Act considerations and jurisdiction-specific requirements
cost-estimation.md	Cost model with real EUR numbers, optimization levers, and scaling inflection points
grafana-dashboard.json	Starter dashboard JSON with Topics Overview, Deep Dive LLM panels, Knowledge Graph, and subscription flow

The Architecture Pattern

Data Sources          Processing           Intelligence         Presentation
┌─────────────┐     ┌──────────────┐     ┌──────────────┐     ┌─────────────┐
│ MS Teams     │────>│ ADF/Airflow  │────>│ Azure OpenAI │────>│ Synapse/    │
│ (Graph API)  │     │              │     │ (GPT-4o/5)   │     │ BigQuery    │
├─────────────┤     │ Nightly batch│     │              │     ├─────────────┤
│ MS Loop      │────>│              │     │ Topic class. │     │ Grafana     │
│ (OneDrive)   │     └──────────────┘     │ Summarization│     │ Tableau     │
└─────────────┘                           └──────────────┘     │ Power BI    │
                                                               │ Any SQL     │
                                                               └─────────────┘

The key design decision: Everything lands in one unified SQL view (MeetingCombinedData). Every downstream tool reads from that single source. No data silos, no sync issues.

How to Use This

1. Start without infrastructure. Do the "Monday Morning" exercise from the talk first — count meetings, ask one lead what they missed, list 10 topics.
2. Pick your stack. The pattern is cloud-agnostic. Swap ADF for Airflow, Synapse for BigQuery, Azure OpenAI for any LLM API.
3. Start with the prompt. The classification prompt is the most immediately reusable artifact. Test it against a few meeting transcripts before building pipelines.
4. Privacy first. Use the checklist before writing any code. Know what you're NOT capturing before you decide what to capture.

Talk Reference Numbers

These are real numbers from a 30-engineer, 3-team org running this system since January 2026:

• 369 meetings processed in ~2 months
• 373 topics classified (60% orphans — the hardest unsolved problem)
• ~2 weeks from idea to first dashboard
• ~EUR 30/mo LLM cost (Azure OpenAI)
• ~EUR 150/mo total infrastructure cost
• $0 Graph API costs (Microsoft removed metered API charges in August 2025)

What's Changed Since the Talk

This starter kit incorporates research and updates through February 2026:

• Graph API: AI Insights API is now GA under /copilot namespace (was beta)
• Graph API: Metered API charges removed (August 2025)
• Teams: Copilot no longer auto-enables persisted transcripts (November 2025)
• Privacy: EU AI Act obligations phasing in; EDPB right-to-erasure enforcement report published
• Privacy: Colorado AI Act (June 2026) and Illinois HB 3773 (January 2026) adding US requirements
• Grafana: Dashboard schema v2 available; Infinity datasource UQL support for LLM response parsing
• ADF: Fabric Data Factory is the strategic direction; migration tools available

License

These templates are provided as-is for educational purposes. Adapt them to your own infrastructure, compliance requirements, and organizational context.