Principal Data Scientist | Causal Inference & Marketing Science | AI / LLM Systems
I lead a team of 8 data scientists. Most of my work is at the intersection of causal inference, marketing mix modeling, and LLM-powered decision support. The hard part in practice is not modeling—it’s turning model outputs into decisions people trust. That’s where I focus.
- Marketing mix modeling and scalable incrementality testing with explicit causal structure
- Turning complex models into decision-ready insights
- Agentic LLM systems built on top of statistical models, not in place of them
- Time series modeling where interpretability is as important as accuracy
SAGE is an AI copilot for interacting with MMMs via natural language. It addresses the insight-synthesis gap that most MMM tools leave open.
- Natural language Q&A over MMM outputs
- Automatic visualizations and diagnostics
- Budget optimization with constrained numerical solvers
- RAG-backed explanations grounded in model artifacts and docs
Cerebro is an agentic system that generates, validates, and executes Bayesian MMM code end-to-end. Agents infer an MMM spec (channels, controls, outcome) from your CSV, then orchestrate six modules—exploration, preprocessing, modeling (NumPyro/JAX), diagnostics, optimization, visualization—with no fixed templates. Hybrid validation (static, API, JAX tracing, execution) with self-healing and optional resume from any stage. Supports modular or monolithic generation and multiple LLM backends (OpenAI-compatible, Ollama, vLLM, MLX).
llm-copilot is an experimental project on LLM-orchestrated analytical workflows around MMMs: tool-calling and orchestration, RAG over model outputs and benchmarks, scoped code execution, and integration with Databricks, Snowflake, BigQuery. Modular by design—where I test ideas before they become products.
Creator and maintainer of DeepCausalMMM, an open-source Python package for MMM that combines deep learning with causal structure. Design principle: interpretability and causal structure over black-box accuracy.
- GRU-based temporal modeling for adstock and lag effects
- DAG-based causal graphs for channel interactions
- Hill-type response curves for saturation and optimization
- Multi-region modeling with shared structure and local effects
pip install deepcausalmmm · Docs
A JOSS paper is under review.
Causal inference: MMM, DAGs, causal discovery (NOTEARS, PC, GES), treatment effects, counterfactual reasoning, budget allocation under uncertainty.
ML & time series: GRU/LSTM, Bayesian and frequentist methods, diagnostics and failure analysis.
LLM systems: Agentic architectures and tool calling, RAG and vector search, evaluation and guardrails for analytical use cases.
Multi-agent LLM systems for analytics, hybrid RAG (structured + unstructured), causal discovery in high-dimensional time series, uncertainty quantification in decision-focused ML, privacy-aware and federated MMMs.
Good models explain why something happened. Great systems help people decide what to do next. That’s the bar I aim for.
Contact
Open to collaboration on LLM systems for analytics, causal inference in production, and open-source data science tooling.