MSc Data Science, London School of Economics and Political Science (LSE)
Right now
I'm finishing an MSc in Data Science at the London School of Economics and working as a research assistant in the Statistics department there.
How I work
I like taking a specific question and testing it properly instead of accepting the surface-level answer. Between statistics, data science, and the tools that come with them, I'm always eager to learn more and put those skills to work on real problems.
Off the clock
Away from all that, I'm usually out hiking, taking photos, or driving somewhere with no fixed plan. I also spend a good chunk of time cooking, playing video games, building LEGO, and listening to music.
London School of Economics and Political Science (LSE)
Coursework in machine learning and data mining, deep learning, artificial intelligence, distributed computing for big data, data analysis and statistical methods, and managing and visualising data. Member of the Data Science Society and the Album Society. Graduating in December 2026.
2020 – 2024
BSc Industrial Engineering
Amirkabir University of Technology (Tehran Polytechnic)
Graduated with a grade of 17.1 out of 20. Coursework in math, linear algebra, differential equations, numerical analysis, probability theory, engineering statistics, optimization, data and information analysis, and management information systems.
2017 – 2020
High School Diploma, Mathematics and Physics
Rabbani Institute of Culture and Education
Graduated with a grade of 19.54 out of 20. Member of the Robotics and Mechanical Engineering Society, building robots and picking up C++ and mechanical design along the way.
Working with Dr. Giulia Livieri and Dr. Chengchun Shi on large language models and Lean 4 for automated theorem proving. The main project formalizes reinforcement learning theory for the open-source Lean Machine Learning (LML) library, which had essentially none. So far it covers the core of planning and learning in Markov decision processes, from Bellman optimality to policy gradients, in 32 machine-checked lemmas and theorems and 12 definitions, with no unproved steps. The sample-complexity theorem needs weaker assumptions than the textbook’s, and the policy gradient proof avoids interchanging derivatives and infinite sums. The work is ongoing, with natural policy gradient and the link between values and expected returns over random trajectories among the next steps, and a possible later stage of using the library to prove a hard, open problem in reinforcement learning theory.
Nov 2025 – Aug 2026
Data Science Researcher
Trilemma Foundation (Capstone) · Vancouver, Canada (remote)
A three-person MSc capstone at LSE, run with the Trilemma Foundation, asking whether a fixed Bitcoin budget can beat plain dollar-cost averaging by putting more money in on favourable days instead of the same amount every day. As a team we built one accumulation model from nine signal groups, drawing on CoinMetrics on-chain history and MicroStrategy’s disclosed purchases alongside Polymarket prediction markets. My part began with the exploratory analysis of more than 2.1 million on-chain and Polymarket records, which made the 365-day drawdown our main on-chain buy signal, and grew into a Polymarket sentiment signal built with a BiLSTM and a fine-tuned FinBERT. Five separate checks showed that signal runs contrarian, though it was left out of the final model because the on-chain signals already carried most of its information. I also designed the MicroStrategy decay kernel and the separate bull and bear regime settings, moved our validation to non-overlapping folds with a holdout we never touched while tuning, and wrote the literature review and results chapters of the report along with its executive summary. After a review with the Trilemma Foundation, the team found look-ahead errors that had pushed an early score up to 79.35%. Once they were fixed, the final model beat plain dollar-cost averaging in 63.6% of 3,075 rolling one-year windows between 2016 and 2025, about nine in ten when markets were flat or falling but fewer than four in ten when they rose.
May 2023 – Jun 2025
Data Analyst
Tanin Print · Tehran, Iran
Translated open business questions from sales and production into defined analytical problems, working with managers to pin down what each question was really asking before choosing data or methods. Integrated multi-source data through SQL Server and built Power BI dashboards to turn the findings into KPI reports, which I presented to non-technical teams in plain language and which improved cross-functional decision-making efficiency by 10%. Also ran exploratory analysis and feature engineering on sales and customer data to build machine learning demand-forecasting models in Python for production scheduling, explaining their assumptions and limits to the people who planned around them, plus a customer engagement system that lifted retention by 20%.
Volunteering
Oct 2025 – May 2026
Technical Consultant
LSESU Data Science and Quantitative Society · London
Mentored three student project teams, nine students in total, on data-driven research through biweekly technical reviews focused on model selection, statistical validation, and reproducibility. Supported projects included a Reddit sentiment analysis for stock prediction, a study of market efficiency in sports betting, and an analysis of market reaction to central bank announcements, and helped translate each into evidence-based reports for non-technical audiences.
Teaching
Jan 2026 – May 2026
Teaching Assistant
K. N. Toosi University of Technology · Tehran, Iran (remote)
TA for the graduate-level Stochastic Processes course under the supervision of Dr. Ardavan Babaei, supporting students on both the theory and the code behind probabilistic models. Prepared lecture materials and programming exercises, and gave detailed feedback on assignments, from the mathematical reasoning to the final implementation.
Sep 2025 – Jan 2026
Teaching Assistant
K. N. Toosi University of Technology · Tehran, Iran (remote)
TA for Entrepreneurship and Business Planning under the supervision of Dr. Ardavan Babaei, supporting students with analytical and business-planning concepts, preparing lecture materials, and giving detailed feedback on assignments to strengthen their reasoning and written presentation.
Co-authored and presented CALM at the conference, an inference-time architecture that attaches a dynamic memory buffer to a large language model to mimic human forgetting and reconstruction bias, reducing overconfidence by 37% and improving human trust scores by 47% in testing.
Revisits the hot hand debate in basketball with two decades of NBA shot data, a hand-written Bayesian Gibbs sampler, and hidden Markov models to separate a genuine effect from statistical illusions and shot selection. The apparent cold hand in the raw data mostly disappears once both are corrected for.
Asks whether a composer's style is a fixed fingerprint or shifts with musical mode, across 295 classical piano works by 19 composers, using hierarchical Bayesian models and hidden Markov models alongside LSTMs and a Transformer for generation. Composer identity explains about a fifth of a piece's structure, but mode barely changes it.
Tests whether joining the EU actually causes faster growth using time series diagnostics, three separate causal inference strategies, and a machine learning model built to recover country-level effects. Once selection bias is stripped out, the average growth effect of membership sits close to zero, though it varies with a country’s distance from the European core.
Processes 5.7 million match events into passing networks across four European leagues to test whether they really play in distinct styles, including a natural experiment on matches played without spectators. Leagues turn out to differ more in where on the pitch teams play than in how their networks are structured.
Combines live weather data from 76 UK cities with 142 years of Met Office records to test whether British weather follows real geographic structure and whether the climate has genuinely shifted. The recent warming trend has clearly picked up speed compared with earlier decades.
Tests whether music taste is statistically linked to anxiety, depression, insomnia, or OCD once age and daily habits are accounted for, cross-checking the strongest result with FDR control over 64 tests, a robust M-estimator and a Bayesian re-analysis. A rock and metal listening factor stayed linked to depression through every check, though the survey design cannot say which causes which.
Mar – May 2026FinanceTime seriesBig data
A Distributed Streaming Pipeline for Real-Time Microstructure Feature Extraction and Volatility Forecasting on the FI-2010 Limit Order Book Benchmark
Built with two teammates for the final project of the Distributed Computing for Big Data course, this streams limit order book data through a distributed pipeline that pulls out real-time microstructure signals on the fly and turns them into short-term volatility forecasts, benchmarked on the FI-2010 dataset.
Built with a team of four for the final project of the Deep Learning course, this models how the S&P 500's implied volatility surface moves across multiple horizons, and teases apart how much of that movement comes from mean-reversion, the surface's own geometry, and broader temporal context.
Builds a distributed pipeline on 29 million US flight records to hunt down what actually predicts a delayed departure, engineering route and time-based features and pitting eight models against each other, from logistic regression to gradient boosting and stacked ensembles.
Dec 2025 – Jan 2026Machine learning
Contextual Bandit Reinforcement Learning for Task Allocation in Restaurant Operations
Built with two teammates for the final project of the Artificial Intelligence course, this uses contextual bandit reinforcement learning to figure out, in real time, who on a restaurant floor should take the next task, learning from context like how busy the floor is rather than following a fixed rulebook.
Oct 2025Machine learning
Mentorship Matching Segmentation: JPMorgan Data for Good Hackathon 2025
Built a segmentation model to match mentors and mentees for JPMorgan Chase's Data for Good hackathon. Worked in a team of five to take it from raw data to a working prototype over a single 12-hour sprint.
Evolved from an initial LSTM-based project forecasting Google's stock prices, this project extends the approach to Bitcoin, comparing ARIMA and LSTM models to examine classical statistical and deep learning methods for financial time series forecasting. Technical indicators, including SMA, EMA, MACD, RSI, and Bollinger Bands, capture market trends and momentum. Using a chronological train-test split and RMSE and MAE for evaluation, the project advances from neural network implementation to comparative out-of-sample assessment of financial forecasting models.
Compares linear, ridge, and gradient-boosted regression models to see which demographic and health factors, such as age, BMI, and smoking status, drive the biggest jumps in medical insurance premiums, and by how much.
Bachelor's thesis using NLP-based sentiment analysis to classify 285,000 Sephora customer reviews as positive or negative with LSTM and BERT, then testing both models on predicting the sentiment of new, unseen comments. Used purchasing-psychology keywords to sort customers into four marketing archetypes, tribe, churner, vampire, and snow leopard, based on their review sentiment and behavior.