About

I am an AI/ML Engineer and NLP Researcher with 2 ACL publications and 5 years of professional experience in machine learning, data science, deep learning, and full-stack software engineering.

My current interests include Language Model Pretraining, Supervised Fine Tuning, Retrieval Augmented Generation (RAG), and Information Retrieval (IR).

My work spans the full stack of these areas, from pretraining language models from scratch, as with a suite of Amharic BERT, RoBERTa, and Llama 3.2 models trained on hundreds of millions of tokens, to fine-tuning and aligning them, as demonstrated by a Phi-2 model refined through supervised fine-tuning and Direct Preference Optimization.

In the retrieval space, I built Splade-Index, an open-source Python package for fast SPLADE-based sparse retrieval, and trained a family of compact SPLADE encoders that outperform BM25 by up to 89%, as well as other SPLADE models up to 6x their size, while running efficiently without a GPU.

My models and tools have been downloaded over 150,000 times on Hugging Face, listed in the Sentence-Transformers documentation, and cited in publications from both academia and industry, including work by Principal Engineers at Amazon Web Services.

Research

• Language model pretraining and supervised fine-tuning
  • Projects: Phi-2 chat models, Minitron chat models, Amharic BERT and RoBERTa, Llama 3.2 Amharic
  • Methods: language model pretraining, supervised fine-tuning, direct preference optimization, tokenizer training, and benchmark evaluation
• Sparse neural information retrieval
  • Projects: SPLADE-Index, SPLADE-Tiny (models), SPLADE-Tiny (code)
  • Focus: sparse neural retrieval, inverted indexes, efficient search, and retrieval-augmented generation
• Selected publications
  • The Multilingual Curse at the Retrieval Layer: Evidence from Amharic — ACL 2026 MeLLM Workshop
  • Optimized Text Embedding Models and Benchmarks for Amharic Passage Retrieval — ACL 2025 Findings