🎯 GTM Attribution Engine

Multi-touch attribution system for B2B SaaS GTM teams. Compares last-touch, linear, Markov chain, and Shapley value models — then uses the Claude API to auto-generate executive narratives, budget reallocation recommendations, and answer natural language questions about channel performance.

---

The problem this solves

Most B2B GTM teams default to last-touch attribution, which systematically over-credits bottom-funnel channels and starves mid-funnel investment. This engine quantifies that distortion and surfaces it in a format an exec can act on.

---

Key features

• 4 attribution models compared side-by-side: last-touch, linear, Markov chain, Shapley value
• Markov chain removal-effect analysis — measures each channel's true incremental contribution
• Shapley value fair-credit allocation — game-theoretically optimal attribution
• Claude API insight generation — auto-writes channel narratives and budget reallocation recommendations
• Natural language Q&A panel — ask freeform questions, get data-grounded answers
• Snowflake-native with CSV demo mode (no credentials required to explore)

---

Architecture

Synthetic Data Generator
        │
        ▼
Snowflake (TOUCHPOINTS table)
        │
        ▼
Python Attribution Models
  ├── Last Touch
  ├── Linear
  ├── Markov Chain (removal effect)
  └── Shapley Value
        │
        ▼
Claude API (claude-sonnet-4-6)
  ├── Channel performance narrative
  ├── Budget reallocation JSON
  └── NL Q&A
        │
        ▼
Streamlit Dashboard (4 tabs)

---

Quickstart

git clone https://github.com/Mohona19/gtm-attribution-engine
cd gtm-attribution-engine

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

cp .env.example .env
# Add your ANTHROPIC_API_KEY (Snowflake credentials optional)

# Generate synthetic data
python data/generate_synthetic_data.py

# Launch dashboard (runs in CSV demo mode without Snowflake)
streamlit run app/streamlit_app.py

---

Project structure

gtm-attribution-engine/
├── data/
│   ├── generate_synthetic_data.py   # Generates 5k journeys, ~18k touchpoints
│   └── sample_data.csv              # 500-row sample for demo mode
├── snowflake/
│   ├── setup_schema.sql             # DDL + views
│   └── load_data.sql                # COPY INTO + validation queries
├── models/
│   ├── attribution_last_touch.py
│   ├── attribution_linear.py
│   ├── attribution_markov.py
│   └── attribution_shapley.py
├── insights/
│   ├── claude_insight_generator.py  # All Claude API calls
│   └── prompt_templates.py          # Prompt constants
├── app/
│   ├── streamlit_app.py             # Main dashboard entry point
│   └── components/                  # Reusable chart & panel components
├── notebooks/
│   └── 01_attribution_exploration.ipynb
├── .env.example
├── requirements.txt
└── DESIGN_NOTES.md                  # Technical decisions & tradeoffs

---

Attribution model comparison

Screenshot of Tab 1 will go here

---

Demo

Loom recording link will go here

---

Design decisions

See DESIGN_NOTES.md for rationale on model selection, LLM prompt architecture, Snowflake schema design, and what I'd build next.

---

Stack

Layer	Technology
Data warehouse	Snowflake
Data generation	Python / NumPy
Attribution models	Python / pandas
AI insight layer	Anthropic Claude API
Dashboard	Streamlit + Plotly
Environment	python-dotenv