Data Science & SciML Researcher · San Francisco Bay Area, CA
About
B.S. in Data Science (Business Analytics track) at Arizona State University, where I conduct faculty-supervised research on the stability and robustness of Graph Neural Networks under topological perturbations. I hold an M.S. in Mathematical Analysis with a concentration in Functional Analysis and Applications of Banach and Nadler Fixed Point Theory. Experienced in production-grade data pipelines, predictive modeling, and large-scale data engineering across complex network infrastructures and scientific machine learning (SciML) domains - where traditional neural networks struggle to capture non-Euclidean dependencies and relational structures. I find structure under complexity and clarity under uncertainty.I bring that theoretical grounding directly into hands-on GNN experimentation in PyTorch Geometric. Explore my featured projects below, or use the navigation icons on the left to dive deeper into my full experience, skills, and background.
Experience
Undergraduate Researcher, GNN Robustness and Stability Analysis — Arizona State University
Conduct faculty-supervised research on the robustness and algorithmic stability of signed and directed Graph Neural Networks under node and edge perturbations, applying Banach and Nadler fixed point theory and contraction stability analysis to large-scale power grid modeling
Design and execute PyTorch Geometric experiments on ASU's Sol supercomputer through SLURM-managed Linux workflows to evaluate convergence behavior and perturbation sensitivity
Build reproducible experimentation pipelines for benchmarking graph model stability under controlled structural changes
Served as founding faculty during Samara University's establishment, leading curriculum development initiatives and serving on multiple ad hoc transition committees to build institutional infrastructure from the ground up.
Taught Linear Algebra, Calculus, and Mathematical Analysis related courses.
Supervised undergraduate thesis research.
Represented faculty interests as elected President of the Samara University Teachers Association across department, college, and University Senate tiers.
High School Math Educator — Local High School
Taught Algebra I & II, AP Calculus A, B/AB, and BC
Developed engaging lesson plans and assessments to enhance student understanding
Collaborated with colleagues to improve curriculum and student outcomes
Mentored students in math competitions and extracurricular activities
Power-system reliability events can carry substantial economic consequences, while engineering teams still face limited time to review large numbers of possible contingencies. GridGuard AI addresses that decision problem by using a scalable PySpark workflow to prioritize modeled post-outage transmission-grid violations on a controlled synthetic IEEE 118-bus DC benchmark.
The project combines scenario-grouped leakage prevention, Spark ML benchmarking, review-budget analysis, uncertainty checks, and an interactive evidence dashboard. At a 20% review budget, Gradient-Boosted Trees captured 95.1% of modeled violations, compared with 78.8% for the strongest tested simple baseline, a 16.3 percentage-point improvement.
Read on to see how GridGuard connects machine learning, power-system reliability, constrained engineering review, and an upcoming energy-market extension involving demand, renewable generation, electricity prices, congestion, and economic-risk prioritization.
These are validation results from a synthetic DC benchmark and do not represent real-grid operational performance or measured utility savings.
Technologies: PySpark, Spark ML, Python, Google Cloud Dataproc, Machine Learning, Power Systems, GitHub
San Francisco Reported-Incident Analytics & Forecasting
Public-safety data is rarely a straightforward measure of safety; it is a complex reflection of reporting behaviors, enforcement shifts, and classification practices. Built on nearly one million SFPD records, this transparency-first time-series model avoids oversimplified point predictions, delivering robust 95% uncertainty intervals and zero-latency forecasting via a deployed Streamlit application.
Read on to see how rigorous rolling-origin evaluation and seasonal decomposition unpacked an 8-year municipal trend without overstating what the data can convey.
Millions in sales mean little when profitability is hidden behind row-level averaging bias and unverified spreadsheets. This project establishes a fully governed six-stage data pipeline that validates $2.30 million in sales, locks down 11 executive key performance indicators, and proves conclusively that discount bands exceeding 20% destroy aggregate margins.
Read on to discover how formula-protected dashboards and granular discount diagnostics transformed reactive reporting into a rigorous operational safeguard.
Technologies: Excel, Power Query, PivotTables, Data Modeling, KPI Governance
When a subscription telecom loses up to a quarter of its subscriber base to silent churn, traditional guesswork or a default 50% threshold leaves money on the table. By deploying a tuned XGBoost engine combined with TreeSHAP attribution and a live Streamlit rescoring suite, this project turns abstract probability into a precision campaign tool—yielding a 3.6 times return on investment on flagged billing exposure.
Read on to explore how precision-recall optimization and custom cost-benefit simulation replaced blind retention spending with targeted, auditable action.
FraudScope: Fraud Detection with Temporal Graph Neural Networks
A production-oriented fraud detection reference that connects chronological machine-learning evaluation, temporal graph research, and a working analyst application. The active model uses calibrated logistic regression, with risk-gated temporal attention running in shadow mode.
Saved synthetic experiments cover 60,000 transactions per seed across seeds 42, 73, and 101. Relational GraphSAGE-style aggregation captured 53.95% of fraud within a 5% review budget, compared with 50.69% for gradient boosting. No GNN won the overall validation-cost comparison.
The application includes transaction evidence, historical account/device links, saved analyst reviews, CSV export, and audit checks. All performance is synthetic. Real IEEE-CIS performance remains unmeasured, and the system is not approved for real financial decisions.
Source: https://github.com/sileshith/fraudscope
Private demo: https://fraudscope-sileshi.bronze-gull-3170.chatgpt.site/ (owner-authorized ChatGPT account required).
Education
B.S. in Data Science (Business Analytics Track) — Arizona State University
Relevant coursework includes Data Mining, Statistical Machine Learning, Database Management, Business Analytics, and Applied Regression Analysis.
Member of the undergraduate research program conducting faculty-supervised research on Graph Neural Network robustness and stability.
Dean's List recipient for three consecutive semesters: Spring 2025, Fall 2025, and Spring 2026.
M.S. in Mathematical Analysis — Bahir Dar University, Ethiopia
Thesis focus: functional analysis, Banach and Nadler fixed point theory, matrix contraction stability, and operator convergence
