Ongoing Research

Active research initiatives, experimental benchmarks, and upcoming preprints across low-resource Semitic NLP, Ethiopic LLM tokenization, and dialect speech recognition.

Here is an overview of my active research projects, ongoing experimental benchmarks, and upcoming papers. My investigations center on bridging language barriers in African & Semitic NLP, Ethiopic LLM tokenization efficiency, clinical RAG architectures, and dialect-robust speech technology.

Clinical RAG Assistant for Maternal & Antenatal Care (ANC)

Active Experimentation
RAG Architecture Hybrid BM25 + FAISS Reciprocal Rank Fusion Clinical Guidelines Offline Edge LLMs

Developing an offline-first, clinically grounded Question Answering system tailored to national Antenatal Care (ANC) guidelines for low-connectivity clinics. Evaluating hybrid retrieval against dense-only baselines to prevent semantic hallucination in critical maternal health contexts.

  • Current Milestone: Implementing Reciprocal Rank Fusion (RRF) combining sparse lexical BM25 with dense bi-encoder embeddings to index bilingual (English & Amharic) medical manuals.
  • Target Models: Local inference using quantized open-weight LLMs (Qwen2.5-0.5B, Gemma 2-2B, Aya) runnable on modest hardware.
  • Evaluation: Hit@k, Mean Reciprocal Rank (MRR), and factual precision verified against clinical case queries.

Overcoming the "Token Tax" in Ethiopic Script LLMs

Empirical Benchmarking
Subword Tokenization Token Tax (9.6x) Byte-Fallback Mitigation Vocabulary Surgery FLORES-200

Investigating cross-lingual token fragmentation where English-centric models require up to 9.6x more tokens to represent equivalent Amharic text, causing context exhaustion and high inference costs.

  • Current Milestone: Designing domain-adapted BPE and Unigram tokenizers trained on Amharic Wikipedia and news corpora with orthographic Fidel normalization.
  • Vocabulary Surgery: Injecting specialized Ge'ez script subwords into base LLM embedding layers while freezing core attention weights to preserve reasoning capabilities.
  • Key Metric: Evaluating Characters-per-Token (CPT), fertility ratio, and cross-entropy reduction on FLORES-200.

Parameter-Efficient ASR for Regional Amharic Dialects

Preprint in Preparation
Speech Recognition (ASR) OpenAI Whisper LoRA / PEFT Wollo Dialect Phonological Adaptation

Standard speech recognition models degrade significantly when exposed to regional Ethiopian dialects. This work applies Low-Rank Adaptation (LoRA) to fine-tune OpenAI Whisper specifically for Wollo Amharic speech patterns and acoustic nuances with low computational overhead.

  • Current Milestone: Acoustic curation and Word Error Rate (WER) / Character Error Rate (CER) benchmarking comparing zero-shot Whisper vs. PEFT adapter checkpoints.
  • Computational Efficiency: Training restricted to low-rank adapter layers on a single GPU while keeping the primary encoder-decoder backbone frozen.

Bits-per-Character (BPC) Evaluation & Retention in Low-Resource LoRA

Analysis & Validation
PEFT Methodology Bits-per-Character (BPC) Bootstrap CIs Catastrophic Forgetting Qwen 2.5

Addressing methodological flaws in standard LLM evaluation where per-token perplexity produces skewed comparisons across tokenizers of varying vocabulary sizes. Demonstrating why Bits-per-Character (BPC) paired with 95% paired bootstrap confidence intervals provides statistically valid measurement of adaptation gains.

  • Current Milestone: Auditing catastrophic forgetting by measuring performance drift on standard English benchmarks (ARC, MMLU) before and after Amharic domain fine-tuning.

Research Collaboration & Compute Partnerships

I welcome collaboration with researchers, academic institutions, and organizations interested in low-resource language technologies, speech recognition for Ethiopian languages, and localized clinical AI.

Feel free to reach out via email at bedruy4@gmail.com or connect on Google Scholar.