Applied Mathematics student at NUST MISIS, focused on Data Science and Machine Learning.

I build end-to-end ML projects — from data validation and leakage-safe evaluation to modeling, interpretation and deployment. My current focus is tabular ML, NLP and practical ML systems.

I am looking for a Data Science or Machine Learning internship or junior position.

Regression project for predicting TVT along hidden segments of horizontal well trajectories.

• Designed leakage-safe validation by complete wells, with a separate spatial guardrail for geographic shift.
• Engineered 32 fold-safe spatial features by reconstructing six geological surfaces with IDW interpolation.
• Combined four CatBoost regressors with distance-aware fallback and boundary correction.
• Improved Public Leaderboard RMSE from 13.608 to 11.670 — a 14.2% reduction.

Python Pandas CatBoost Feature Engineering Group CV Spatial Validation

Binary classification and ranking project for prioritizing clients under a 10% outbound-call budget.

• Removed post-call data leakage and used temporal folds to evaluate performance on future campaigns.
• Compared linear and tree-based models, selecting Logistic Regression for its stability across time.
• Achieved AP 0.530, ROC-AUC 0.745 and Lift@10% 1.91 on the final period.
• Explained the ranking with permutation importance and model coefficients.

Python Pandas Scikit-learn Temporal Validation Ranking Metrics Model Interpretation

Multiclass NLP project for routing customer requests across 77 banking intents.

• Built a shared leakage-safe evaluation pipeline with fixed splits and macro F1.
• Compared TF-IDF with Logistic Regression, a custom Transformer Encoder implemented in PyTorch and fine-tuned DistilBERT.
• DistilBERT reached 0.882 validation macro F1 and 0.886 official test macro F1; TF-IDF remained a strong baseline at 0.861.

Python Scikit-learn PyTorch NLP Transformers DistilBERT

Personalized recommendation system built on RetailRocket e-commerce interaction data.

• Generated Top-10 product recommendations from users’ views, cart additions and purchases.
• Compared Popularity, Item-KNN, ALS and hybrid approaches using a 14-day temporal holdout.
• Selected binary Item-KNN, achieving NDCG@10 of 0.0456 compared with 0.0033 for the Popularity baseline.
• Evaluated recommendation quality separately for cold, light and active users using ranking and catalog coverage metrics.

Python Pandas Recommender Systems Item-KNN Ranking Metrics Temporal Validation

GPT-style language model trained from scratch for controllable short-story generation.

• Implemented a 126M-parameter decoder-only Transformer in PyTorch with RoPE, GQA, RMSNorm, SwiGLU and KV cache.
• Built the byte-level BPE tokenizer, data pipeline and mixed-precision training loop with gradient accumulation, validation and checkpointing.
• Added a FastAPI web application for steering generation with a summary, required words, story opening and narrative features.

Python PyTorch Transformer Tokenizers FastAPI Generative NLP

AI Observability, FinOps and Governance platform for operating AI products and agents.

• Built an AI registry and telemetry pipeline for agent runs, LLM calls, tool calls and business outcomes.
• Implemented token-cost accounting, budgets, ROI and waste metrics, policy violations, audit history and role-based access.
• Exposed a REST ingestion API and Python SDK, with FastAPI, PostgreSQL, Redis and React services packaged through Docker Compose.

FastAPI PostgreSQL Redis React Docker Compose REST API