AI / ML Engineer & Researcher

Bethel Melesse
Tessema

Founder of PARAHIGG, an AI-powered legal research platform for Ethiopian law, with an MSc in Artificial Intelligence and a background in AI/ML engineering and research.

Portrait of Bethel Melesse Tessema

About Me

A bit about me

I’m an AI/ML engineer and researcher with a master’s in Artificial Intelligence and the founder of PARAHIGG.

My work sits at the intersection of deep learning, language, and science. Whether it’s building data and retrieval models for low-resource languages, applying graph neural networks to drug discovery, or now making Ethiopian law searchable, I’m drawn to problems where AI can reach people and fields that technology has overlooked.

My path took me from software engineering studies in Ethiopia to AI research in South Korea. Along the way, I volunteered with NGOs and served as a translator and interpreter for the Korean War Veterans Revisit Program. Those experiences shape how I approach technology: as something that should connect people, lower barriers, and create practical value.

3+ Years in AI/ML
4.31 M.S. CGPA / 4.5
1000+ TB Data Processed

Experience

What I've worked on

Founder

Aug 2026 — Present

PARAHIGG — Seoul, South Korea
parahigg.com

  • AI-Powered Legal Research for Ethiopian Law: Founded and lead PARAHIGG, a platform that makes Ethiopian legislation and case law searchable and easier to understand for lawyers, students, and the public.
  • Product & Engineering: Building the platform end to end, from collecting and processing legal documents to retrieval and LLM-based question answering over Ethiopian law.

AI Engineer

Oct 2024 — Jul 2026

AI Team, 3Billion — Seoul, South Korea

  • Forward Synthesis Evaluation Model: Developed a neural network to score synthetic feasibility of reaction pathways in an RL-based molecular generation pipeline. Optimized training with PyTorch, RDKit, and DeepChem, achieving a 24× speedup (days → 1 hour). Benchmarked with AUROC, AUPRC, and Tanimoto similarity alongside experimental chemists.
  • End-to-End Protein-Ligand Interaction Prediction: Built an E2E model predicting binding pockets, ligand poses, and binding affinities for structure-based drug discovery. Led data preprocessing and graph construction from PDB/MOL2/SDF formats; designed multi-task GNN architectures with attention; implemented distributed multi-GPU training with MLflow tracking.

AI Researcher

Mar 2022 — Feb 2024

DBDC Lab, Ajou University — Suwon, South Korea

  • Information Retrieval for Low-Resource Languages: Trained a BERT-like transformer for information retrieval in African languages, benchmarking against TF-IDF and BM25.
  • Generalized Output Spaces for Classifiers: Improved cross-domain and cross-task generalization in multi-modal classifiers by fusing class/task descriptions, achieving state-of-the-art zero-shot performance on CIFAR, RCV1, and 20NG.
  • Low-Resource Datasets for LLM Adaptation: Processed 1000+ TB of Common Crawl data for under $2, creating monolingual datasets for 7 low-resource languages, and fine-tuned multilingual LLMs with QLoRA on consumer GPUs.

Education

Foundations of my craft

2022 – 2024

M.S. in Artificial Intelligence

Ajou University, Suwon, South Korea

CGPA: 4.31 / 4.5

2018 – 2022

B.A. in International Studies

Hankuk University of Foreign Studies, Seoul, South Korea

CGPA: 3.75 / 4.5

2015 – 2017

Software Engineering

Addis Ababa Institute of Technology, Ethiopia

3 semesters

Publications & Projects

Research and projects

Production System

Production-Grade RAG: PARAHIGG

A legal research assistant for Ethiopian federal law in English and Amharic, covering Federal Supreme Court cassation decisions and Ministry of Justice proclamations, regulations, and directives. Answers are grounded only in the retrieved law, with numbered citations that link back to the source documents.

Preprint

UnifiedCrawl: Aggregated Common Crawl for Affordable Adaptation of LLMs on Low-Resource Languages

A low-cost pipeline that extracts and deduplicates low-resource language text from all of Common Crawl, producing far larger monolingual datasets. Fine-tuning multilingual LLMs on it with QLoRA improves Amharic perplexity and few-shot scores on consumer hardware. Also published in the KICS Proceedings.

Personal Project

Deep Learning Models, from Scratch

Built foundational deep learning models from first principles in PyTorch and NumPy: MLPs, CNNs, RNNs, a full encoder–decoder Transformer with attention and positional encoding, a decoder-only language model, and a tokenizer. The work also spans Computer Vision and Graph Neural Networks.

Skills

Tools of the trade

Domains Legal information retrieval · Low-resource languages · Drug discovery · ML fundamentals

RAG & Retrieval

PARAHIGG legal search; low-resource IR research at Ajou.

RAG pipelines Hybrid search (dense + BM25) Reciprocal rank fusion Cross-encoder reranking Multi-query retrieval Metadata filtering Dense retrieval (BERT-style) TF-IDF / BM25 baselines

LLMs & Prompting

PARAHIGG question answering; UnifiedCrawl fine-tuning.

Grounded generation with citations Query rewriting Conversational memory Prompt versioning LLM cost tracking QLoRA fine-tuning OpenRouter Ollama Haystack Hugging Face Transformers sentence-transformers

Multilingual NLP

English and Amharic legal text; UnifiedCrawl datasets.

English ↔ Amharic translation Amharic text normalization Legacy-font-to-Unicode conversion Lexicon-based decoding Low-resource language datasets Multilingual LLM adaptation

Drug Discovery

Forward synthesis and protein–ligand models at 3Billion.

Protein–ligand interaction Binding affinity & pose prediction Synthetic feasibility scoring RL-based molecular generation PDB / MOL2 / SDF processing RDKit DeepChem

ML Fundamentals

M.S. in AI and every project since.

Supervised & unsupervised learning Classification & regression Feature engineering Train / val / test splits Cross-validation Overfitting & regularization Gradient-based optimization Hyperparameter tuning Loss functions & metrics

Deep Learning

GNNs at 3Billion; Transformers and CNNs built from scratch.

Transformers from scratch Graph Neural Networks Multi-task learning Attention mechanisms Zero-shot & multi-modal classification CNNs / RNNs Computer Vision Distributed multi-GPU training

Data & Documents

PARAHIGG legal corpus; 1000+ TB of Common Crawl.

Web crawling (BeautifulSoup, REST APIs) PDF processing (PyMuPDF, pikepdf, fontTools) OCR (Tesseract) Docling layout extraction Structure-aware chunking Contextual embeddings Resumable batch pipelines Common Crawl at scale (1000+ TB) Deduplication

Evaluation & MLOps

Retrieval metrics, chemistry benchmarks, LLM evals.

Hit rate / MRR / precision AUROC / AUPRC Tanimoto similarity Perplexity & few-shot evals Pipeline tracing MLflow

Databases & Storage

PARAHIGG production stack.

Qdrant (vector DB) PostgreSQL (Supabase) Cloudflare R2 Supabase Storage

Languages & Tools

Daily drivers across all work.

Python Java Bash PyTorch (CUDA) TensorFlow NumPy Docker uv Conda Git / GitHub Linux pytest mypy Ruff

Off the Clock

What I do for fun

Hiking

Hiking

1,573 mHighest summit

Scuba Diving

Scuba Diving

50+Dives
65 minLongest dive
10.7Best SAC (L/min)

Festivals & Concerts

Festivals & Concerts

  • Guns N’ Roses
  • Deep Purple
  • Bryan Adams
  • Dua Lipa

Working Out

Working Out

  • Calisthenics
  • Weight training
  • Pilates

TV Shows

TV Shows

  • Grey’s Anatomy
  • Criminal Minds
  • Early-2000s shows

Music

Music