About
I’m a data scientist and applied ML researcher with 6 years of end-to-end project ownership across experimentation, causal inference, applied ML, and LLM systems. I’ve designed publication-grade randomized studies with 1,232 participants, shipped multi-agent AI applications, and translated statistical findings into decision-ready recommendations.
My work spans both research and engineering: running large behavioral experiments and formal analytical models on the research side, and building agentic AI systems, RAG pipelines, and data infrastructure on the engineering side. I’m focused on applied scientist, data scientist, AI engineer, and operations research roles where rigorous causal, ML, and engineering skills improve real-world data and product decisions.
Technical Skills
- Statistics, ML & Causal Methods
- Python, SQL, scikit-learn, TensorFlow, logistic / OLS / Poisson regression, A/B testing, causal inference, instrumental variables, difference-in-differences, hypothesis testing, cross-validation, feature engineering, exploratory text analysis, group-level fairness metrics.
- Agentic AI & LLMs
- OpenAI Agents SDK, CrewAI, LangChain, Pydantic structured outputs, multi-agent orchestration, RAG, prompt engineering, RAG evaluation.
- Engineering & Decision Science
- FastAPI, PostgreSQL, Gradio, oTree, Heroku/Railway, Git, Gurobi, LaTeX, Excel, R, Optimization, Simulation, Forecasting, Queuing, Decision Analysis, Bayesian Decision Modeling.
Professional Experience
Doctoral Researcher
Haskayne School of Business, University of Calgary · Calgary, AB
- Designed and ran two randomized behavioral experiments with 1,232 participants, testing how fairness norms, accountability, and accuracy targets affect ML model-selection decisions.
- Built the full experimental stack in oTree, deployed on Heroku, recruited through Prolific, and analyzed results with logistic regression and exploratory text analysis.
- Found fairness-norm messaging nearly doubled the odds of submitting the fairest model; high accuracy targets raised the odds of submitting the least fair model by up to 1.71.
- Developed a Bayesian and rational-inattention model of human-AI decision-making, proving threshold conditions under which group-specific AI error disclosure improves accuracy and fairness.
Instructor
Haskayne School of Business, University of Calgary · Calgary, AB
- Taught 250+ undergraduates across 5 sections of Business Analytics and Logistics Management, with ratings of 5/5 in the most recent term and 7/7 in earlier terms.
- Translated optimization, simulation, forecasting, queuing, and decision analysis into applied managerial decision frameworks for non-technical audiences.
Industrial Engineering Intern
Ezam Automotive Parts Group (PJS) · Karaj, Iran
- Analyzed strategic management workflows, identified process inefficiencies, and built a web-based prototype to automate targeted processes within the internship period.
Projects
CrewAnalyst
A 7-agent data-analysis pipeline built on CrewAI. Upload a CSV, get back a markdown and PDF report with statistics, anomalies, correlations, visualizations, and an executive narrative. Parallel async execution, Pydantic-typed schema contracts between agents, and model-tier routing for cost-latency tradeoffs.
- Architected a 7-agent pipeline (profiler, statistician, anomaly, correlation, visualizer, synthesizer, reporter) that converts an uploaded CSV into a polished markdown and PDF analysis report with descriptive statistics, anomaly flags, correlation insights, charts, and an executive narrative.
- Engineered parallel async execution via CrewAI's async_execution task flag for 3 independent analytics agents (statistician, anomaly, correlation), cutting analysis time before the visualization step. Used Pydantic-typed schema contracts as explicit handoff interfaces between agents, and model-tier routing (Haiku for structural/classification tasks, Sonnet for reasoning and synthesis) for cost-latency tradeoffs.
See a report this agentic workflow created
GitHub repoDude
A 3-agent natural-language interface (OpenAI Agents SDK) for a budgeting app. Translates plain-English questions about spending into structured retrieval and analytics over transaction data. Deployed in a FastAPI / PostgreSQL / React app in daily production use.
- Built a 3-agent system (conversational, retrieval, analyst) that translates plain-English budgeting questions into structured retrieval plus analytics over transaction data.
- Deployed into a FastAPI / PostgreSQL / React app in daily production use, integrated via Claude Code from an isolated mock-data prototype to the live database.
GitHub repoRagTeX
A LaTeX-aware RAG pipeline over structured academic papers. Uses cleaned .tex sources, section-aware chunking, FAISS embeddings, and metadata-filtered retrieval — extended with naive vs. advanced RAG comparisons, an LLM router for query-specific retrieval strategies, and a LaTeX parser for equations, headers, and appendices.
- Built a RAG pipeline over structured LaTeX academic papers using section-aware chunking, hybrid FAISS + BM25 retrieval, MMR reranking, and metadata-filtered search.
- Implemented an LLM router that selects between naive and advanced retrieval strategies based on query type, and built a LaTeX parser to cleanly extract equations, section headers, and appendix content.
GitHub repoResearch
Shaping Programmer Practices: Mitigating Bias in ML Development.
Designed and ran two randomized behavioral experiments (N=604 and N=628; 1,232 participants total) simulating an end-to-end ML hiring-model pipeline under varied fairness and accountability interventions. Reject & Resubmit, Production and Operations Management Journal.
- Built the full experimental stack: interface in oTree, deployment on Heroku, recruitment via Prolific, analysis with logistic regression, Wald chi-square tests, and exploratory text analysis.
- Found fairness-norm messaging nearly doubled the odds of submitting the fairest model (OR approx. 2.0); high accuracy targets raised the odds of submitting the least fair model (OR up to 1.71); the combined norm + accountability intervention showed a negative interaction (OR=0.43), revealing non-additive effects with direct governance implications.
Collaborative Fairness: Human and Machine Interaction.
Developed a formal Bayesian + rational-inattention model of human-AI decision-making; proved threshold conditions under which disclosing group-specific AI error rates improves both aggregate accuracy and demographic fairness. Working paper. Presented at MSOM, CORS, and AIMOR 2025.
- Identified a non-obvious failure mode in which awareness can reverse disparity direction under specific parameter regions, showing why aggregate accuracy audits can mask group-level harm.
Education
Ph.D. in Operations and Supply Chain Management
University of Calgary, Haskayne School of Business · Calgary, AB
M.Sc. in Industrial Engineering
Sharif University of Technology · Tehran, Iran
B.Sc. in Industrial Engineering
Kharazmi University · Tehran, Iran
Awards & Service
- Paramount Resources Ltd Graduate Scholarship in Business, University of Calgary (2024, 2025).
- Runner-up, Outstanding Achievement in Teaching Award, University of Calgary (2025).
- Treasurer & Board Member, Alberta Student Chapter of the Canadian Operational Research Society (2024-2025); Session Chair, CORS Annual Conference (2024).