About
I build machine learning systems that ship, and I do research on speech and language technology for languages that most models ignore. Over four years at Ishango.ai I deployed production LLM systems for clients in Germany, Canada and the UK: a RAG search agent on AWS, a recommendation system, a document QA assistant over 2 TB of content, and the data pipelines underneath them. I'm now a master's student in Data Science at Duke University.
My research is on text-to-speech and speech recognition for low-resource languages, LLM post-training and instruction tuning, and multilingual evaluation. I have published at IEEE ASRU and IJCNLP-AACL on multilingual evaluation and low-resource speech and language technology, and I have trained and deployed TTS and ASR models for low-resource languages.
Resume
Mila – Quebec AI Institute
Research Intern
May 2026 – Aug. 2026- Built a face-conditioned emotional text-to-speech system in PyTorch on the MEAD corpus, conditioning synthesis on facial expression to control emotion.
- Reached 72% emotion accuracy on unseen speakers, 93% cross-speaker voice transfer, and 6.9% WER on the generated speech.
Ishango.ai
Data Scientist / Data Engineer
Oct. 2022 – Aug. 2026- Deployed a RAG search agent on AWS for a mental-health platform, replacing tag-only search with natural-language retrieval and improving search efficiency by 80%.
- Shipped five LLM systems for an advisory firm: an article recommendation system, a document chatbot, a QA and search assistant over 2+ TB of Drive content, a Neo4j stakeholder knowledge graph, and a pipeline scraping 10,000+ media articles.
- Co-developed a Python data-synchronization framework with checkpointing, change detection, retries and scheduling, so a new data source is onboarded as a small, uniform job.
- Built and operated production ingestion jobs publishing thousands of market, weather and price time series into a trading database, with CI/CD and monitoring across dev, preprod and prod; owned failure diagnosis, backfills and data verification.
Indigenius AI
Speech and Language Processing Researcher (part-time)
Jul. 2025 – Dec. 2025- Trained TTS and ASR models for low-resource Nigerian languages.
Duke University
Master's in Data Science
- Relevant courses: Modeling and Representation of Data, Data Engineering Systems, Natural Language Processing, Deep Learning.
African Institute for Mathematical Sciences (AIMS)
MSc in Mathematical Sciences, Data Science
- One-year intensive master's program. Grade: Distinction.
- Thesis: Deep Learning Methods for Text-to-Speech with Application to the Ethiopian Ge'ez Language.
- Relevant courses: Data-driven Optimization with ML Applications, High Performance Computing, Statistical Regression, Big Data Analytics with Python, Database and Data Management, Principles of Statistical Learning, Operations Research.
Mekelle University - Mekelle Institute of Technology
BSc in Computer Science and Engineering
- Grade: Honors (3.77/4.0)
- Project: developed a rule-based stemmer for Tigrinya to support downstream NLP tasks.
- Internship: iCog Labs, Addis Ababa (Sep. 2019 – Jan. 2020).
- Relevant courses: Data Structures, Design and Analysis of Algorithms, Operating Systems, Microcontrollers, Compiler Design, Microprocessors, Cryptography and Network Security, Digital Signal Processing.
ML / AI
PyTorch, Hugging Face, LLM fine-tuning (SFT, LoRA), RAG, FAISS and vector search, TTS / ASR, scikit-learn, Neo4j
Data Engineering & MLOps
AWS (ECS, SageMaker), GCP (GCS, Vertex AI, Cloud Run), Docker, Prefect, Terraform, FastAPI, Azure DevOps CI/CD, Datadog, Git, Linux
Languages
Python, SQL, R
- Teaching Assistant, AIMS Rwanda (2022–2023): Introduction to Algorithms, Python and LaTeX; Advanced Excel for Data Analysis for the Evidence-Based Climate Reporting Initiative.
- Co-founder, TigraiCodes (2019–2020): programming and computer science for high-school students in Mekelle.
- Teaching Assistant, AddisCoder (2019) and Iga Coding Python Camp (2022).
CIFAR Inclusive AI Scholarship
Awarded to attend the CIFAR Deep Learning and Reinforcement Learning (DLRL) Summer School.
See the NISR Statistical Dashboard under My Work.
My Work
Publications, models and projects. Most are about making speech and language technology work for languages with little data.
Ge'ez Text-to-Speech
End-to-end TTS for Ge'ez, a liturgical language with no prior speech corpus. Built the text normalization and alignment pipeline, and trained Tacotron 2 + WaveGlow. Master's thesis at AIMS: Deep Learning Methods for Text-to-Speech with Application to the Ethiopian Ge'ez Language.
Tigrinya GPT-2 Model
51M-parameter GPT-2 model trained from scratch on 21M Tigrinya tokens. Built to advance text generation for a low-resource African language with limited NLP tooling.
Gemma-3-4B-Tigrinya
Google's Gemma 3 (4B) fine-tuned via SFT on curated Tigrinya data for fluent text generation in a language with minimal LLM support.
Gemma-3-4B-Tigrinya QA
Instruction-tuned variant of Gemma 3 (4B) optimized for Tigrinya question answering, enabling conversational AI capabilities for low-resource language speakers.
NISR Statistical Dashboard
1st PlaceInteractive Python dashboard for exploring Rwanda's statistical datasets (GDP, CPI, trade). Simplified complex economic indicators into intuitive visual narratives for public accessibility.
Technical Articles
-
Implicit Euler method for Ordinary Differential Equations (ODEs) using Python
A detailed explanation of the Implicit Euler method for solving ODEs with Python implementation
-
Explicit Euler method for Ordinary Differential Equations (ODEs) using Python
Implementation and explanation of the Explicit Euler method for solving differential equations
-
Trapezoidal Methods for Solving ODE Systems
Implicit and predictor–corrector trapezoidal schemes with Python examples.
-
Runge–Kutta Methods: 2nd- and 4th-Order Schemes
Mid-point (RK2) and classic RK4 methods explained with examples.
Publications
-
mSTEB: Massively Multilingual Evaluation of LLMs on Speech and Text Tasks
IEEE ASRU 2025 · First author
A benchmark evaluating large language models on speech and text tasks across many languages, with a focus on how performance degrades for low-resource ones.
mSTEB Leaderboard -
Ibom NLP: A Step Toward Inclusive Natural Language Processing for Nigeria's Minority Languages
IJCNLP-AACL 2025
Datasets and baselines for underrepresented Nigerian languages.
-
OpenBibleTTS: Large-Scale Speech Resources and TTS Models for Low-Resource Languages
Under review
Large-scale aligned speech corpora and open TTS models for low-resource languages.
Visual Diary
I have a passion for capturing moments through my lens, especially those that showcase the beauty of nature. Take a look...
Contact
Open to summer 2027 internships in ML, speech, and applied AI, and to collaborations on low-resource language technology.