Writing

Independent research paper, in review, 2026. What quantization does to Amharic small language models on a consumer GPU: 4-bit costs about 3 points, 3-bit about 20 or more, and tokenizer fertility predicts the damage.

Why one-shot vector search falls short on complex questions, and how agentic, multi-hop retrieval plans its searches, iterates and refines queries before it answers.

Predicting whether a student drops out, stays enrolled or graduates: exploratory analysis, PCA, an XGBoost model and a Streamlit app for live predictions.

Notes from a Kaggle forecasting challenge: predicting monthly and quarterly Walmart sales by product category from historical sales data, starting with data cleaning.

Analyzing and forecasting Brent crude prices: data exploration, cleaning, feature engineering, first models, and a move to Markov-switching ARIMA.

Fraud detection for e-commerce and credit card transactions, end to end: preprocessing, model training, explainability, and deployment behind an API.

A warehouse for data on medical businesses scraped from public Telegram channels: Telethon scraping, cleaning and dbt transformations, YOLO object detection on images, and a FastAPI layer on top.

Building a credit scoring system: credit-risk concepts, exploratory analysis, feature engineering, and weight of evidence and information value for model building.

Marketing analytics from public data: Telegram channel posts and Google Play reviews and downloads loaded into PostgreSQL, then cleaned, explored and tested statistically.