Sales Call Transcript Extractor

Transform sales call transcripts into actionable CSV data with AI-powered sentiment analysis, business insights, and deal prediction.

Built with LaunchDarkly AI Configs and Vercel AI Gateway for dynamic schema management. Uses the official @launchdarkly/server-sdk-ai-vercel SDK with intelligent tool selection. Includes ML model training for deal prediction using CatBoost.

What You Get

📊 40-65 structured data fields per call including sentiment scores, deal signals, engagement metrics, and call-specific insights - all exportable as CSV for your CRM or analytics tools.

---

Quick Start

Manual Setup

# 1. Install dependencies
npm install
python3 -m venv venv && source venv/bin/activate
pip install catboost scikit-learn pandas numpy joblib requests python-dotenv

# 2. Configure environment variables
cp .env.example .env
# Add your keys to .env:
#   LAUNCHDARKLY_SDK_KEY=sdk-xxxxx
#   LD_API_KEY=api-xxxxx (for bootstrap only)
#   LD_PROJECT_KEY=your-project (for bootstrap only)

# 3. Get Vercel AI Gateway token
npx vercel env pull

# 4. Bootstrap LaunchDarkly AI Config (one-time)
python bootstrap/create_unified_config.py

# 5. Run
npm run dev  # → http://localhost:3000

Docker Compose

# 1. Configure environment
cp .env.example .env  # Edit with your keys
npx vercel env pull   # Get OIDC token

# 2. Start services
docker-compose up

# 3. Setup LaunchDarkly (one-time, in another terminal)
docker-compose exec python sh -c ". /app/venv/bin/activate && python bootstrap/create_unified_config.py"

# → http://localhost:3000

---

How It Works

Unified AI Extraction

Upload Transcripts (.txt or .md)
    ↓
AI analyzes content and selects appropriate schema
    ↓
Extracts 40-65 fields in single LLM call
    ↓
CSV downloads automatically

6 Extraction Schemas

The AI automatically selects the best schema based on transcript content:

Schema	Type	Fields	Use Case
A	Prospecting	~43	Cold outreach, first contact, gatekeeper conversations
B	Discovery	~48	BANT qualification, needs assessment, pain point identification
C	Demo	~58	Product demonstrations, feature walkthroughs, POC discussions
D	Proposal	~53	Pricing negotiations, contract terms, commercial discussions
E	Technical	~63	Architecture reviews, integration planning, technical deep-dives
F	Customer Success	~53	QBRs, renewal discussions, expansion opportunities

Core Fields (in all schemas):

• Identity: transcript_id, customer_company_name, salesperson_name, call_date
• Business: deal_stage, customer_size, industry, estimated_deal_value
• Sentiment: overall_sentiment_score, sentiment_about_product, sentiment_about_pricing (-1 to +1)
• Engagement: customer_engagement_score, urgency_score, budget_confidence_score (0 to 1)
• Statistics: transcript_word_count, customer_question_count, competitor_mention_count
• Signals: next_steps_defined, competitors_mentioned, decision_makers_present

---

Configuration

1. Bootstrap LaunchDarkly

export LD_API_KEY="api-xxxxx"
export LD_PROJECT_KEY="your-project"
python bootstrap/create_unified_config.py

What it creates:

• ✅ 1 AI Config: transcript-extraction-unified
• ✅ 6 extraction tools (schemas A-F)
• ✅ 1 unified variation with all tools attached
• ✅ Default model: claude-3-7-sonnet-latest

Important: The script deletes all existing AI configs and tools to ensure a clean state.

2. Environment Variables

# LaunchDarkly
LAUNCHDARKLY_SDK_KEY=sdk-xxxxx       # Runtime SDK key
LD_API_KEY=api-xxxxx                 # For bootstrap script only
LD_PROJECT_KEY=your-project          # For bootstrap script only

# Vercel AI Gateway (choose one)
VERCEL_OIDC_TOKEN=eyJhbGc...         # Preferred - auto-refreshed
AI_GATEWAY_API_KEY=vck_xxxxx         # Alternative - manual

Get OIDC token:

npx vercel env pull  # Refresh every 12 hours

---

Usage

Feature Extraction

1. Start dev server: npm run dev
2. Upload .txt or .md transcripts
3. Click "Extract Features"
4. Watch real-time progress
5. CSV downloads automatically

Features:

• Real-time progress tracking with SSE
• Per-file error reporting (continues on failures)
• Empty file detection
• UTF-8 and special character support
• Large file handling

ML Model Training

1. Click "Train Model" in the ML section
2. Choose demo data or use extracted CSV
3. View metrics: Precision, Recall, F1, RMSE, R²
4. Model saved as deal_model.pkl

---

Deployment

Local Development

npm run dev  # → http://localhost:3000

Production (Vercel)

npx vercel --prod

Vercel Setup:

1. Add LAUNCHDARKLY_SDK_KEY in project settings
2. VERCEL_OIDC_TOKEN auto-provided in production
3. Build configured with .npmrc and next.config.js

Production URL: https://vercel-tan-beta-71.vercel.app

---

Dynamic Configuration

All schemas live in LaunchDarkly - update without code changes!

Modify Schemas

1. LaunchDarkly → AI Tools → Select tool
2. Edit JSON Schema parameters
3. Save → Changes apply immediately

A/B Testing

1. Create new variation with modified schema
2. Set targeting rules (e.g., customer_segment = "Enterprise")
3. Monitor metrics in LaunchDarkly dashboard

Benefits:

• No code deployment to change fields
• Gradual rollout and canary testing
• Performance tracking per variation
• Instant rollback capability

---

Customization

Add New Schema

1. Define schema in LAUNCHDARKLY_TOOLS.json:

{
  "variation_g_custom": {
    "function": {
      "name": "extract_custom_features",
      "description": "Extract features from custom call type",
      "parameters": {
        "type": "object",
        "properties": {
          "custom_field_1": { "type": "string" },
          "custom_field_2": { "type": "number" }
        },
        "required": ["custom_field_1"]
      }
    }
  }
}

2. Update bootstrap in bootstrap/create_unified_config.py:

tool_mapping = {
    # ... existing tools ...
    "extract_custom_features": ("Extract Custom Features (G)", "variation_g_custom"),
}

3. Run bootstrap:

python bootstrap/create_unified_config.py

Adjust Model Selection

Change model in bootstrap/create_unified_config.py:

model_config_key="Anthropic.claude-3-7-sonnet-latest"  # Premium, accurate
# or
model_config_key="Gemini.gemini-2.0-flash"  # Fast, cheaper

---

Project Structure

├── app/                              # Next.js app
│   ├── page.tsx                      # Upload UI
│   └── api/
│       ├── extract-stream/           # SSE endpoint
│       ├── extract/                  # Non-streaming endpoint
│       └── train-model/              # ML training endpoint
├── lib/
│   ├── pipeline.ts                   # Unified extraction pipeline
│   └── launchdarkly-client.ts        # SDK integration with tool calling
├── bootstrap/
│   ├── create_unified_config.py      # Setup unified AI config
│   └── create_configs.py             # Legacy (do not use)
├── ml/
│   └── train_and_return_metrics.py   # CatBoost model training
├── LAUNCHDARKLY_TOOLS.json           # Schema definitions
├── docker-compose.yml
└── .env

---

Troubleshooting

"OIDC token has expired" (401 errors)

npx vercel env pull  # Refreshes token
# Restart dev server

"Unified extraction config not found"

python bootstrap/create_unified_config.py
# Check LAUNCHDARKLY_SDK_KEY is set

"No transcripts were successfully processed"

• Check OIDC token is fresh
• Verify LaunchDarkly config exists in dashboard
• Check console for specific errors

LaunchDarkly connection issues

# Verify SDK key
echo $LAUNCHDARKLY_SDK_KEY

# Test initialization
npm run dev
# Look for "LaunchDarkly client initialized successfully"

Build errors

• Already configured in next.config.js
• Uses webpack externals for LaunchDarkly SDK
• Check .npmrc has legacy-peer-deps=true

---

API Reference

POST /api/extract-stream

Streams extraction progress via Server-Sent Events.

Request:

{
  "transcripts": [
    { "name": "call1.txt", "content": "..." },
    { "name": "call2.txt", "content": "..." }
  ]
}

Response (SSE):

event: progress
data: {"type":"extraction","current":1,"total":2,"percentage":50}

event: complete
data: {"type":"done","csvData":"...","results":[...]}

POST /api/extract

Non-streaming extraction (for testing).

Returns: { success: true, csvData: "...", results: [...] }

---