Principal Data Scientist | Causal Inference & Marketing Science | AI / LLM Systems

I lead a team of 8 data scientists building models and systems that actually get used. Most of my work sits at the intersection of causal inference, marketing mix modeling, and LLM-powered decision support systems.

A recurring theme in my work:

the hard part isn’t modeling — it’s turning model outputs into decisions people trust.

That’s where I focus.

• Marketing Mix Modeling and Scalable Incrementality Testing Systems with explicit causal assumptions
• Translating complex models into decision-ready insights
• Agentic LLM systems that sit on top of statistical models (not instead of them)
• Time series modeling where interpretability matters as much as accuracy

SAGE is an AI copilot designed to help users interact with MMMs using natural language.

Most MMM tools fail at insight synthesis. SAGE was built to close that gap — even if it means being opinionated about workflows.

What it does well:

• Natural language Q&A over MMM outputs
• Automatic generation of visualizations and diagnostics
• Budget optimization using constrained numerical solvers
• RAG-backed explanations grounded in model artifacts and documentation

🔗 Live demo: https://SAGEinsights.streamlit.app 🔗 Repo: https://github.com/adityapt/SAGE

llm-copilot is a more experimental project exploring how LLMs can orchestrate analytical workflows around MMMs.

It focuses on:

• Tool-calling and orchestration (LLMs as controllers, not predictors)
• Retrieval-augmented generation over model outputs and benchmarks
• Safe, scoped code execution for exploratory analysis
• Integration with existing data platforms (Databricks, Snowflake, BigQuery, etc.)

This project is intentionally modular — it’s where I test ideas before they harden into products.

I’m the creator and maintainer of DeepCausalMMM, an open-source Python package for MMM that combines deep learning with causal structure.

Design philosophy:

Prefer interpretability and causal structure over black-box accuracy.

Key ideas:

• GRU-based temporal modeling for adstock and lag effects
• DAG based causal graphs to capture channel interactions
• Hill-type response curves for saturation and optimization
• Multi-region modeling with shared structure and local effects

📦 pip install deepcausalmmm 📖 Docs: https://deepcausalmmm.readthedocs.io

A JOSS paper is currently under review.

Causal Inference

• MMM, DAGs, causal discovery (NOTEARS, PC, GES)
• Treatment effects, counterfactual reasoning
• Budget allocation under uncertainty

ML & Time Series

• GRU / LSTM models
• Bayesian and frequentist approaches
• Model diagnostics and failure analysis

LLM Systems

• Agentic architectures and tool calling
• RAG systems and vector search
• Evaluation and guardrails for analytical use cases

Academic profile: ORCID — https://orcid.org/0009-0008-9495-3932

• Multi-agent LLM systems for analytical workflows
• Hybrid RAG (structured + unstructured retrieval)
• 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 try to hold myself to.

• LinkedIn: https://www.linkedin.com/in/adityapt/
• GitHub: https://github.com/adityapt
• Email: puttaparthy.aditya@gmail.com

I’m always happy to collaborate on:

• LLM systems for analytics
• Causal inference in real business settings
• Open-source data science tooling