Matiwos Desalegn

• Built an enterprise multi-agent RAG document assistant - LangChain + CrewAI over Llama 3.3 70B (Groq), ChromaDB hybrid retrieval with cross-encoder reranking, Keycloak SSO, and PostgreSQL conversation memory
• Built the end-to-end credit-scoring ML pipeline - Feast feature store (S3 registry, Redis online, Aurora offline) with automated feature-refresh across batch domains; FastAPI inference on EKS with <200ms p95 latency
• Built an AI data-standardization agent using a provider-agnostic LLM layer (OpenAI / Ollama / HuggingFace), containerized and deployed on AWS
• Developed a FastAPI embedding service for Ethiopian-language retrieval (EthioLLM; Amharic/Ge'ez/Afaan Oromo/Tigrinya/Somali) with batch processing and a fallback model

• Built the organization's first AWS-managed lakehouse (Athena + Glue Data Catalog) orchestrated with Mage AI - 6-hour micro-batch ingestion for institutional partners, 300+ tables / 7,000+ columns, high-watermark extraction, and date partitioning
• Designed a lightweight real-time CDC pipeline under compute constraints - OLake for MySQL binlog to Apache Iceberg (Polaris catalog), queryable via Trino
• Built and deployed the Batch API microservice (FastAPI + EKS) ingesting domain payloads from partner institutions; persists to MongoDB and pushes features to Feast (Redis) for credit-scoring consumption
• Built production Firebase Analytics export pipelines - BigQuery to S3 Bronze Parquet to Iceberg Silver via PyIceberg, with partner-isolated Airflow DAGs and Secrets Manager + IRSA

• Designed and implemented the frontend of an enterprise HR/payroll (ERP) system, improving usability and operational efficiency
• Integrated RESTful APIs with backend teams for reliable, end-to-end data flow
• Achieved a 40% reduction in payroll processing time through optimized data handling and a responsive interface

• Intensive 12-week data science and machine learning fellowship program
• Worked on real-world data science projects with industry mentors
• Specialized in MLOps, feature engineering, and model deployment
• Collaborated in agile teams on end-to-end ML projects

• Led a community of 200+ developer students in learning cutting-edge technologies
• Organized workshops and events on cloud computing, AI/ML, and web development
• Mentored students in building real-world projects and technical skills
• Facilitated hackathons and coding competitions

Full-stack OCR application for recognizing Ethiopic/Amharic script from images and PDFs, powered by the HHD-Ethiopic pre-trained deep learning model with TensorFlow + CTC decoding, hierarchical text detection (lines → words → characters), and an interactive Next.js detection canvas with confidence scoring

TensorFlow inference with CTC decoding on 306-char Ethiopic alphabet · Three-level hierarchical detection: line → word → character

Multilingual agricultural AI platform supporting Amharic, Afan Oromo, and English for Ethiopian farmers

300ms average response time · 85% accuracy in recommendations

Fintech-focused AI solution for credit risk assessment and financial planning

95% credit scoring accuracy · 60% faster loan processing

Full-stack AI-native Revenue Cycle Management platform for the GCC healthcare market — transforms the traditional 10-step RCM workflow into an intelligent, automated system powered by Google Gemini 2.0 Flash, with a patient-centric data model, automated appeal generation, and Zendesk integration

Async FastAPI with Gemini 2.0 Flash inference · Targets 30% revenue leakage reduction via denial prevention

Cloud-native FastAPI pipeline that ingests insurance claim data from heterogeneous EMR sources (CSV + JSON), normalizes schemas, and applies hybrid deterministic + Google Gemini AI classification to flag claims as retryable or non-retryable for resubmission

100% test scenario pass rate across all 6 test categories · Hybrid deterministic + AI eliminates unnecessary LLM calls

Generalizable AI agent that executes user-defined tasks across multiple SaaS platforms (Zendesk, Freshdesk) using natural language commands and Playwright browser automation — with session persistence, provider abstraction, and LLM-powered intent extraction

Headless Playwright execution with session reuse · Single natural language command drives multi-platform automation

10 Academy · Remote

Intensive 12-week program — graduated top 10% of cohort. Specialized in MLOps, feature engineering, and end-to-end ML project deployment.

Wolaita Sodo University · Wolaita Sodo, Ethiopia

Grew community from 50 to 200+ members; organized 12 workshops and 3 hackathons with 100+ participants each.