Eighteen years of turning models into decisions people use.
I am a hands-on technical lead: I architect the system, write production code, set the evaluation bar and work directly with executives on what to build next. Most recently that has meant agentic AI on regulated data, and forecasting and optimization platforms that run planning across hundreds of stores.
Hands-on technical lead for a portfolio of concurrent ML and generative-AI initiatives, working directly with director and executive leadership and mentoring junior and senior data scientists and interns.
lower forecast error
~50%
Retail demand forecasting platform, ~350 stores, vs. the legacy model
markdown revenue
+11%
Markdown optimization, most recent benchmark
high-confidence coding
99.76%
51,518 card transactions classified by an agentic workflow on regulated PII
compute cost
−90%
Assortment space optimization rebuild (runtime 9 h → 75 min)
Agentic AIRAG & retrievalEvaluation & guardrails
Work Order Intelligence Platform
AI Solution Architect & Lead AI Engineer
The problem
Maintenance work orders arrive as free text. Diagnostics, asset details and costs were buried in prose and typed in by hand, so nobody could trust or analyze them at scale.
What I built and led
Architected LangGraph agentic workflows that extract Problem–Cause–Remedy diagnostics from HVAC and refrigeration work orders, with agentic RAG over BigQuery vector search to match assets and components to the asset taxonomy and spec sheets.
Built an LLM-as-judge evaluation pipeline that surfaced upstream data-quality issues and drove data-capture fixes at the source.
Classified every work order into category-specific cost buckets: a JEV model (Typesafe AI) as the primary classifier, an LLM fallback below 0.9 probability, and a five-criteria evaluation step that ships, reviews or rejects each result.
Designed a proof of concept for a multimodal (text and image) intake agent that replaces form-based requests: it gathers missing details, triages, and hands off each request with similar past incidents, likely duplicates and repair documents.
Corporate-card transactions were coded by hand against an inconsistent category list, on data containing regulated personal information.
What I built and led
Architected and led a LangGraph agentic workflow on Vertex AI that discovered a governed 44-category taxonomy and then classified every transaction against it.
Designed for regulated data from day one: governed categories, confidence scoring, and automatic OpenAI-to-Gemini failover so the pipeline keeps running when a provider does not.
Planners needed SKU-by-store forecasts for regular and promoted items far enough ahead to plan buys and promotions, and the legacy model was both less accurate and slow.
What I built and led
Led a cross-functional team of data scientists and engineers, setting technical direction and MLOps standards across 22 repositories.
Built the production models: LightGBM for weekly SKU-store sales and promotional redemption (benchmarked against XGBoost and a Temporal Fusion Transformer) and ARIMA+ for daily demand, trained on Ray in Vertex AI, served via BigQuery ML and fed by Dataform feature pipelines.
Designed the evaluation and explainability layer — leakage-audited backtests by category, promotion and horizon week, plus SHAP dashboards for planners — and an experimentation platform.
Delivered an on-demand Cloud Run forecasting API returning per-promotion forecasts in about 25–30 seconds.
Earlier, led modeling for the legacy promotional forecasting platform this replaced, and helped refactor it from R and Scala to PySpark on AWS, cutting its runtime by nearly 70%.
Markdowns are a trade-off between revenue, shrink and empty shelves, and they were set by rule of thumb across three retail channels.
What I built and led
Led the price-elasticity component on Databricks/PySpark: an ensemble of XGBoost, LightGBM and AdaBoost with monotonic constraints, trained on four years of store-SKU POS history to produce 38-week demand curves.
Led the optimization engine: a pruned combinatorial search over multi-week discount schedules that balances revenue against shrink and empty-shelf cost, with channel-specific weights, hard inventory caps and merchant guardrails.
Delivered weekly recommendations that merchants typically adopt as-is, with overrides in the UI.
Every store has a fixed amount of aisle space. Deciding how wide each product section should be — including widths a store has never tried — is a forecasting and an optimization problem at once.
What I built and led
Led a two-stage hierarchical XGBoost model (segment to store, monotonic constraints) that forecasts response curves for every candidate width.
Formulated a PuLP mixed-integer optimizer that reallocates each store's space to the highest-return sections under merchandising constraints, solved in parallel across stores on Databricks/Spark.
Delivered it through a self-service Streamlit dashboard, and rebuilt the preprocessing, feature engineering and compute layers.
Fulfillment centers need to know, hour by hour and weeks ahead, how many orders and units each store pairing will generate in order to staff and plan capacity.
What I built and led
Architected a 21-day multi-horizon forecast of hourly orders and units for every EFC–store pair (9 EFCs, 73 stores), ensembling an ARIMA baseline, global LightGBM models and a Temporal Fusion Transformer.
Engineered the MLOps pipeline on Databricks (AWS): MLflow-tracked experimentation and model selection, weekly retraining and daily inference via model serving.
Feeds EFC staffing and capacity planning, with staff allocation partly live.
Distribution centers plan labor and trucks weeks in advance; forecasts had to be accurate across the full horizon and monitored continuously.
What I built and led
Architected and productionized a Databricks/Spark system on AWS forecasting outbound and pallet selection for 12 distribution centers over a 28-day horizon, with per-series XGBoost models, MLflow tracking, weekly retraining and daily scoring.
Built Spark-native accuracy monitoring over 2.2 million forecast–actual pairs, and shipped v2 as a zero-downtime shadow deployment.
Built PipeLearner, an internal Spark-distributed training package with Optuna tuning, adopted across data-science teams and cutting model training and deployment time by more than half.
Teach three programs — two certificate programs from Johns Hopkins University and a post-graduate program from UT Austin's McCombs School of Business — covering production LLM applications end to end, from prompt optimization and retrieval to multi-agent systems, evaluation and guardrails.
students taught
250+
instructor rating
4.92 / 5
university programs
3
Agentic AIRAG & retrievalEvaluation & guardrails
Certificate Program in Agentic AI
Johns Hopkins University
The program
Building single- and multi-agent systems in Python, from LLM and retrieval foundations to agents that plan, use tools, coordinate and can be evaluated and governed.
Main topics I teach
How LLMs work, prompting techniques and model selection
RAG pipelines: chunking, embeddings, vector stores, hybrid search and reranking
Evaluating GenAI systems and reducing hallucinations
Single-agent systems: ReAct, tools, memory and MCP
Multi-agent architectures and agent-to-agent (A2A) coordination
Agent evaluation: task success, reasoning trajectories and adversarial tests
Responsible AI: guardrails, human-in-the-loop review and audit trails
Post Graduate Program in AI Agents for Business Applications
UT Austin McCombs
The program
AI agents for business teams: designing agents that plan, use tools and remember, applying them across finance, HR, retail, healthcare and customer service, and judging which use cases are worth automating.
Main topics I teach
Generative AI and LLMs for business, prompt techniques and reusable templates
RAG and agentic RAG for context-aware answers
Agents with tools and short- and long-term memory, via MCP
Planning, reasoning, self-reflection and task decomposition (ReAct)
Designing multi-agent systems
Testing and evaluating agents, with human-in-the-loop review
Securing and deploying agents: privacy, access control and compliance
Certificate Program in Applied Generative AI and Agentic AI
Johns Hopkins University
The program
Production LLM applications end to end: prompting and its optimization, retrieval, fine-tuning, multi-agent systems, and knowing which approach fits a given problem.
Main topics I teach
Prompt engineering: few-shot, chain- and tree-of-thought, self-consistency, ReAct
Prompt stability and programmatic prompt optimization
Evaluation with LLM-as-a-judge and gold-standard datasets
Embeddings, RAG and GraphRAG
Fine-tuning small language models — and when to prefer prompting or RAG
Serve on the advisory board of a recovery-science wellness startup, advising the founding team on technology strategy, platform architecture and AI/ML strategy across the product.
Syne pairs a nightly recovery practice with an app that turns a member's own wearable data (Oura, Whoop or Apple Health) into one clear action each evening.
Research Assistant · Instructor · Teaching Assistant
Doctoral research on the security of cyber-physical manufacturing systems, alongside teaching.
Developed quantitative risk models using game theory (Nash and quantal-response equilibrium) and Monte Carlo simulation to produce risk scores, loss distributions and control ROI.
Applied optimization to shop-floor scheduling in a virtual manufacturing network, and statistical and ML methods to educational research on student success.
2013 – 2015 · Esfahan, Iran
Reef Chemical Industries Complex
Manufacturing / Enterprise Data Analyst
Developed, validated and deployed an ML-based anomaly-detection system for predictive maintenance using supervised and unsupervised methods.
Reduced downtime by 30% and saved more than $200K a year.
2009 – 2013 · Esfahan, Iran
Isfahan Steel Complex
Data Analyst / Consultant
Led the implementation of an optimized maintenance system, using statistical and ML models to forecast machinery health and reduce unplanned downtime and emergency repairs.
Foundations
Education & skills
2019
Ph.D., Mechanical & Manufacturing Systems Engineering (Cybersecurity)
University of Texas at San Antonio, Klesse College of Engineering