Featured Projects
Production AI systems built for scale and reliability
Latest Publications
Research contributions to the AI community
A Noise-Robust End-to-End Framework for Amharic Speech Recognition
Research Square (Under Review - International Journal of Speech Technology) · 2025
End-to-end automatic speech recognition (ASR) framework tailored specifically to the Amharic language. Our approach integrates a convolutional neural network (CNN), a recurrent neural network (RNN), and Connectionist Temporal Classification (CTC) to directly transcribe speech into text—bypassing the need for labor-intensive dictionary creation. We evaluate our method on a large corpus of 20,000 noisy Amharic utterances, achieving a word error rate (WER) of just 7%. This result underlines the effectiveness of our system in handling challenging acoustic conditions. By reducing complexity and manual overhead, our end-to-end model offers a practical and accurate solution for real-world deployments, with broader implications for developing ASR in other low-resource and noise-prone environments.
Epileptic Seizure Detection Using Deep CNNs with Spectral and Complexity Features from EEG Signals
Journal Publication · 2025
We present a clinically oriented deep learning approach for automated detection of epileptic seizures from scalp electroencephalography (EEG) recordings. The proposed method integrates domain-specific features, including spectral band-power and signal complexity measures such as Shannon entropy and Hjorth parameters, into a compact one-dimensional convolutional neural network (CNN). We evaluate the model on the Children's Hospital Boston and Massachusetts Institute of Technology (CHB-MIT) pediatric EEG dataset. While the model achieves high discrimination, with an area under the ROC curve of about 0.98 and accuracy of about 99% on held-out balanced data, the default decision threshold of 0.5 yields poor seizure recall. By lowering the threshold to 0.10 based on validation analysis, seizure sensitivity improves to about 91%, with a manageable false-alarm rate of about two to three per hour. Beyond accuracy, we report cross-validation with confidence intervals, event-based detection metrics, and ablation studies comparing spectral, complexity, and raw EEG inputs.
Develop the Herbs Plant Classification Model for Patient Care Using Deep Learning
Master's Thesis - Bahir Dar University, Faculty of Computing · 2023
This research focuses on noise removal in the classification and description of herbal plants using the Xception model. The study aims to improve accuracy by addressing specific challenges such as shadow removal, dust removal, and color correction through the integration of advanced algorithms. The research incorporates the Shadow Detection and Removal (SDR) algorithm for shadow removal, the Top-hat transform algorithm for dust removal, and the Gray World Algorithm along with Contrast Limited Adaptive Histogram Equalization (CLAHE) for color correction and histogram enhancement. The evaluation demonstrates exceptional accuracy, achieving a classification accuracy of 99% and validation accuracy of 97.8%.
Professional Experience
Current roles and key achievements
AI Engineer & Team Lead
CurrentAmplitude Ventures AS
Norway · Oct 2024 - Present
AI Engineer & Data Scientist
Zare Innovations
Addis Ababa, Ethiopia · Jan 2024 - Sep 2024
AI Engineer & Data Scientist
Amhara Public Health Institute
Bahir Dar, Ethiopia · Feb 2023 - Dec 2023
Tech Stack
Tools and technologies I work with