CV

The Geometry of Dormant Defection (Latent Sleeper Agent Linear Probes)

Mechanistic interpretability research and 28-layer probing replication suite investigating layer-wise linear probe dynamics for detecting latent sleeper agents and dormant backdoors in large language models. GitHub Code Research Article Zenodo Preprint

• Layer-Wise Activation Probing: Designed and evaluated linear probes across all 28 transformer layers of Qwen2.5-Coder-1.5B-Instruct to trace where deceptive policies linearly separate from aligned representations.
• Mid-Layer Concentration Discovery: Discovered that peak linear separability concentrates in mid-layer representations, formulating trigger-agnostic defensive monitors that catch backdoors without inspecting outputs.
• Tech Stack: Python, PyTorch, Hugging Face Transformers, Linear Probes, Mechanistic Interpretability, Qwen2.5-Coder.

Amharic Hate Speech Detection

Deep learning NLP research developed as B.Sc. Capstone at Addis Ababa University, utilizing fine-tuned multilingual BERT (mBERT) to detect online hate speech and abusive language in low-resource Ethiopic NLP. Hugging Face Model GitHub Code Jupyter Notebook

• Research & Model Weights: Fine-tuned multilingual BERT (mBERT) representations capturing Amharic morphology; published trained weights and tokenizers on Hugging Face Model Hub.
• Evaluation Pipeline: Established robust evaluation benchmarks for classifying hate, offensive, and neutral content on low-resource social discourse datasets.
• Tech Stack: Python, PyTorch, Hugging Face Transformers, Datasets, mBERT, Jupyter/Colab.

AfroChat

AI-powered multi-model enterprise chat platform bringing cutting-edge LLMs together with custom workspace solutions, streaming architectures, and multi-tenant security. Live App

• Distributed Backend: Engineered real-time streaming chat completions and workspace orchestration using WebSockets, Redis pub/sub, and RabbitMQ message queues.
• Production Cloud Infra: Containerized services with Docker and Kubernetes, backed by MongoDB clusters and AWS S3/SES for assets and automated transactional mail.
• Tech Stack: Express, Node.js, MongoDB, OpenAI API, AWS S3/SES, Redis, WebSockets, Docker, Kubernetes, RabbitMQ.

Adot

Comprehensive maternal healthcare platform supporting pregnant women through personalized tracking, expert doctor consultations, and predictive AI health monitoring (consulted by MIT alumnus Adis Ojeda). Live Platform

• Fullstack & Mobile Ecosystem: Architected web client and companion mobile interface delivering clinical appointment booking and personalized trimester milestone tracking.
• Predictive Health Modeling: Implemented predictive analytics for prenatal risk factors with secure cloud storage, HIPAA-conscious design, and real-time alerts.
• Tech Stack: Next.js, Node.js, MongoDB, Flutter, Django, TypeScript, Tailwind CSS, GCP, Firebase, Docker.

Cursor Spend Tracker

Telemetry and cost analytics SaaS platform tracking Claude token spend and BYOK charges with automated Vercel cron background data synchronization. Live App GitHub Code

• Telemetry & Cost Modeling: Automated cost breakdown across Claude models (Sonnet, Opus, Haiku), cached input token ratios, and live burn-rate analytics.
• Serverless Architecture: Built with Next.js 16 App Router, Prisma ORM, Neon Serverless Postgres, and automated Vercel Cron jobs.
• Tech Stack: Next.js 16, TypeScript, Tailwind CSS, Prisma, Neon Postgres, Recharts, Vercel.

promptKit

Automated prompt engineering, synthetic test-case generation, and LLM evaluation benchmarking toolkit for generative AI applications. GitHub Code

• Systematic Benchmarking: Generates diverse edge-case test distributions and benchmarks prompt candidate performance using automated LLM-as-a-judge scoring.
• Interactive Dashboard: Streamlit-based interactive exploration suite for comparing model outputs, latency, and token efficiency.
• Tech Stack: Python, Streamlit, OpenAI API, Prompt Tuning, LLM Evaluation.

VitalAI

Django-based predictive healthcare REST API delivering personalized health insights and anomaly alerts from Apple Health telemetry data. GitHub Code

• Biometric Ingestion Pipeline: Ingests, parses, and normalizes Apple HealthKit metrics (resting heart rate, active energy, sleep cycles) into standardized timeseries.
• Predictive Health Modeling: Machine learning algorithms identifying health metric anomalies and providing actionable lifestyle recommendations.
• Tech Stack: Python, Django REST Framework, Docker, Apple HealthKit API, NumPy, Pandas.

Redash Chatbot Add-on

Conversational Business Intelligence extension for Redash, enabling natural language questions, automated SQL query generation, and live chart visualization. GitHub Code

• Text-to-SQL Engine: Translates complex plain-text analytical questions into validated SQL queries against underlying data warehouses.
• Redash Integration: Integrates directly with Redash API to execute generated queries and render interactive visualizations dynamically.
• Tech Stack: Python, LangChain, Redash API, LLM text-to-SQL, PostgreSQL.

RateEat

Community food discovery and restaurant review platform featuring personalized recommendations, interactive menu browsing, and verified user ratings. Live App

• Recommendation Engine: Implemented smart collaborative filtering for dish discovery tailored to user taste preferences and dietary habits.
• Fullstack Application: Responsive Next.js/React frontend with Tailwind CSS and an Express/MongoDB backend with Firebase real-time notifications.
• Tech Stack: React, Next.js, Tailwind CSS, Node.js, Express, MongoDB, Redux, Firebase.

Co-founder and CTO of Vula.

Co-founder and CEO of Vula.

Professor and Department Chair, EECS at UC Berkeley; Founder of AddisCoder.

MIT Alumni; Founder & CEO of Theseus Group; Consulted Adot.

Associate Professor and Dean of ECSE at Addis Ababa University; Former Academic Professor.

CEO of A2SV (Africa to Silicon Valley).