How saikumarljmu-maker/agentic-marketing-ai works
This repository implements retrospective Criteo campaign-budget decision simulations, with rule-based, LLM-assisted, and portfolio-strategy variants, plus tools for counterfactual and statistical evaluation. In the sampled LLM loop, daily observations are analyzed and turned into campaign recommendations; the loop compares those with human decisions and can project KPI effects. The V2 preparation and analysis modules document a separate portfolio-replay capability. The README is empty and several modules were not sampled, so uncertain connections are omitted.
Main entry points: simulation_loop_llm.py LLM replay · simulation_loop.py Rule-based replay · observer.py Observation · response_curves.py Response curves
Components 20
Simulation workflows
- Rule-based replay
src/simulation_loop.py - LLM replay
src/simulation_loop_llm.py - V2 replay
src/simulation_loop_v2.py
Decision agents
- Observation
src/agents/observer.py - Analyst
src/agents/analyst.py - Campaign strategy
src/agents/strategy.py - Portfolio strategy
src/agents/portfolio_strategy.py - Decision memory
src/agents/memory.py - Decision comparison
src/agents/execution.py
Data preparation
- Criteo preparation
src/data/criteo_prep.py - Processed campaign data
src/data/criteo_prep.py
Evaluation
- Response curves
src/evaluation/response_curves.py - Replay analysis
src/evaluation/analyse_v2.py - Statistical analysis
src/evaluation/statistical_analysis.py - Protection replay
src/evaluation/protect_off_replay.py - Model comparison
src/model_comparison.py
Other components
- Researcher
- Criteo dataset
- Anthropic API
- Ollama models
Connections 30
- Researcher → Rule-based replay: runs replay
- Researcher → LLM replay: runs replay
- Researcher → V2 replay: runs replay
- Criteo dataset → Observation: reads dataset
- Rule-based replay → Observation: initializes
- LLM replay → Observation: initializes
- Observation → Rule-based replay: returns observations
- Observation → LLM replay: returns observations
- Rule-based replay → Decision memory: retrieves context
- Rule-based replay → Decision comparison: records decisions
- Rule-based replay → Response curves: fits curves
- Rule-based replay → Response curves: projects effects
- LLM replay → Decision memory: retrieves context
- LLM replay → Analyst: requests analysis
- LLM replay → Campaign strategy: requests decisions
- LLM replay → Decision comparison: records decisions
- LLM replay → Response curves: fits curves
- LLM replay → Response curves: projects effects
- Analyst → Anthropic API: requests analysis
- Campaign strategy → Anthropic API: requests decisions
- Criteo preparation → Criteo dataset: reads dataset
- Criteo preparation → Processed campaign data: writes aggregates
- V2 replay → Portfolio strategy: requests allocation
- Portfolio strategy → Anthropic API: requests strategy
- Processed campaign data → Replay analysis: provides replay data
- Processed campaign data → Protection replay: provides replay data
- Statistical analysis → Statistical analysis: computes comparisons
- Researcher → Model comparison: compares models
- Model comparison → Anthropic API: requests completion
- Model comparison → Ollama models: runs model