Cync Summaries

Cync Summaries uses a multi-agent AI system to generate executive summaries of team activity.

How It Works

┌─────────────────────────────────────────────────────────┐
│                   Multi-Agent Pipeline                   │
├───────────┬───────────┬──────────┬───────────┬─────────┤
│ Collector │ Historian │  Editor  │ Validator │Polisher │
│  (fetch)  │ (context) │ (write)  │  (score)  │(refine) │
└─────┬─────┴─────┬─────┴────┬─────┴─────┬─────┴────┬────┘
      │           │          │           │          │
      ▼           ▼          ▼           ▼          ▼
   Data      Historical   Draft      Quality    Final
  Sources    Context     Summary     Scores    Summary

Agent Roles

Agent	Role	Output
Collector	Fetches data from all sources	Raw facts
Historian	Adds historical context	Enriched context
Editor	Writes initial summary	Draft document
Strategist	Identifies strategic implications	Business insights
Validator	Scores quality (5 dimensions)	Quality metrics
Polisher	Final refinement	Executive summary

Quality Scoring

The Validator agent scores summaries on 5 dimensions:

Dimension	Weight	Description
Specificity	20%	Contains concrete details
Accuracy	25%	Facts are correct
Density	20%	Information per word
Outcomes	20%	Focuses on results
Context	15%	Provides background

Threshold: Weighted score must be 8.0+ to pass.

Usage

Generating a Summary

1. Navigate to /protected/cync
2. Click "Generate Last Week Summary"
3. Watch agents process in real-time
4. Review quality scores
5. Read final executive summary

Agent Theater UI

The UI shows:

• Agent status (queued/active/complete/failed)
• Progress indicators
• Live thought stream
• Source data card (after Collector)
• Quality score panel (during Validator)

Data Sources

The Collector agent fetches from:

1. Wrike - Tasks created, updated, completed
2. Fireflies - Meeting transcripts
3. Stickies - Recent notes
4. News - Articles saved

Configuration

AI Model

# .env.local
OPENROUTER_API_KEY=sk-or-v1-xxx
OPENROUTER_MODEL=anthropic/claude-3.5-sonnet

Supported Models

Free Models (testing):

• meta-llama/llama-3.2-3b-instruct:free
• google/gemma-2-9b-it:free

Paid Models (production):

• anthropic/claude-3.5-sonnet
• openai/gpt-4o

API

Generate Summary

GET /api/generate

Returns Server-Sent Events (SSE) stream with agent updates:

{
  "agent": "collector",
  "status": "active",
  "progress": 50,
  "message": "Fetching Wrike tasks...",
  "data": { "facts": [...] }
}

Event Types

Status	Description
queued	Agent waiting
started	Agent beginning
active	Agent processing
complete	Agent finished
failed	Agent error (may retry)

Retry Logic

If Validator fails (score below 8.0):

1. Editor receives feedback
2. Editor revises summary
3. Validator re-scores
4. Process repeats (max 3 attempts)