Portfolio - Yoseph Negash Kibe

Hello, I am

Yoseph Negash Kibe

Data Aspirant

About Me.

Hello! I'm Yoseph, a data scientist and AI/ML practitioner who enjoys building impactful, data-driven solutions from machine learning and deep learning to NLP systems and production-ready web applications.

Tools & Technologies:
🐍 Python for building and training ML and deep learning models end-to-end.
🗄 SQL and Databricks for data engineering and scalable pipelines.
📊 Matplotlib, Seaborn, Streamlit, FastAPI, and Gradio for visualization and deployment.

Areas of Expertise:
📈 Machine learning and predictive models across finance, healthcare, and e-commerce.
🤖 Computer vision and NLP systems with real-world deployment.
📉 End-to-end ML pipelines from raw data to production.

I'm committed to lifelong learning, hands-on projects, and connecting with others who believe in using AI and data to make a difference. Always open to collaborations, let's build something meaningful together.

1

Python Project

3

Machine Learning Projects

2

Deep Learning Projects

Skills.

Machine learning

Python

SQL

Deep Learning

NLP

Computer Vision

Projects.

My Experience.

Technical Instructor | FDRE Technical and Vocational Training Institute, Addis Ababa
Jan 2024 – Present

Delivered programming instruction across Python, C, Data Structures & Algorithms, databases, and web development
Guided students in algorithmic thinking, software design principles, and clean code practices
Mentored learners in debugging, system design, and practical problem-solving approaches

IT Intern | Ethiopian Engineering Corporation, Addis Ababa
Sep 2022 – Nov 2022

Configured SharePoint sites and streamlined document management workflows across departments
Resolved network connectivity and hardware/software issues across supported workstations
Supported user access management and system administration tasks for internal users

Awards & Certificate.

Python: Beginner to Advanced For Data Professionals

SQL for Data Science

Math and Statistics For AI, Data Science

Master Machine Learning for Data Science & AI: Beginner to Advanced

Deep Learning: Beginner to Advanced

Natural Language Processing

Applied Biomedical AI

Contact.

Feel free to get in touch with me. I am always open to discussing new projects, creative ideas or opportunities to be part of your visions.

Download CV

[email protected]

https://www.linkedin.com/in/yosephnegash/

GitHub

https://github.com/joss413

Deep Learning Project

Ethiopian Cuisine Classification + Bilingual NLP Assistant

Domain/Function: Computer Vision + Natural Language Processing

An end-to-end AI system that classifies 10 Ethiopian food dishes from images using an ensemble of DenseNet121 and ResNet50 achieving 98% accuracy, with a bilingual NLP assistant that answers questions in English and Amharic.

View Project
Engagement

View Project
on GitHUB

View Project
on Gradio

Ethiopian Cuisine Classification + Bilingual NLP Assistant

Ethiopian Cuisine AI app homepage showing the upload interface, Classify Food button, and supported foods footer including Beyaynetu, Doro Wat, Kitfo and 7 more Ethiopian dishes. Built with Gradio.

App predicting Beyaynetu with 99.66% confidence and answering questions in both English and Amharic using NLP. Demonstrates bilingual Q&A with TF-IDF cosine similarity matching.

Project Details

This project combines deep learning and NLP to classify Ethiopian food images and answer questions about them in English or Amharic.

The deep learning component uses an ensemble of DenseNet121 and ResNet50 trained with Optuna hyperparameter tuning, achieving 98% validation accuracy across 10 Ethiopian food classes. Class imbalance was handled using weighted loss functions.

The NLP component uses TF-IDF vectorization and cosine similarity to match user questions to the correct topic, with automatic language detection using Unicode range detection for Amharic script support.

The system is deployed as a Gradio web application with a FastAPI backend, allowing users to upload a food image, get an instant prediction, and ask questions about the predicted food in either English or Amharic.

Tech Stack: PyTorch, DenseNet121, ResNet50, Optuna, TF-IDF, Gradio, FastAPI, Python

Project video

AI-Powered Daily Content Newsletter Automation using n8n

Domain/Function: Data Engineering / AI Content Automation

Built an n8n workflow that runs every morning at 8 AM, fetches content from YouTube, Medium, and Dev.to via RSS feeds, scores each item using an AI agent, and delivers the top 5 ranked articles as a styled HTML email newsletter.

View Project
Engagement

Full n8n workflow: Schedule Trigger → RSS (YouTube, Medium, Dev.to) → Merge → Edit Fields → 24hr Filter → AI Agent (OpenRouter, Structured Output Parser) → Sort → Google Sheets → Limit → Aggregate → HTML → Gmail.

Auto-generated HTML email newsletter titled "Data, AI & Analytics Digest" delivered to Gmail inbox, displaying AI-curated top content on Data Analytics, GenAI, Data Science, and Automation.

Project Details

This project automates daily content curation using an AI-powered n8n workflow. Every morning at 8 AM, a schedule trigger fetches new posts from YouTube, Medium, and Dev.to via RSS feeds and merges them into a single pool. The data is cleaned using Edit Fields, then filtered to only include content published in the last 24 hours. An AI Agent powered by Google Gemini evaluates each item against topics like Data Analytics, GenAI, and Automation, assigning a unique relevance score from 0 to 10. Results are sorted by score, and all items scoring 5 or above are logged to Google Sheets as a permanent archive. The top 6 items are then aggregated and formatted into a professional HTML email with score badges, clickable titles, source info, and AI-generated descriptions. Finally, Gmail delivers the newsletter automatically to the inbox. This project demonstrates skills in RSS integration, AI prompt engineering, structured output parsing, HTML templating, and end-to-end workflow automation.

Project video

Automated Variance Report Generation using n8n AI Agent

Domain/Function: Data Engineering / Data Quality Automation

Built an n8n AI agentic workflow that fetches daily CSV data from email, compares it against PostgreSQL database records using Gemini AI, generates a variance report in Google Docs, and emails the report link back automatically.

View Project
Engagement

Automated Variance Report Generation using n8n AI Agent

Full n8n agentic workflow: Gmail Trigger → Extract CSV → Aggregate → Edit Fields → Variance Report Generator (Gemini, Memory, Postgres, Google Docs, Calculator, Drive) → Reply to Email.

Automated Gmail reply containing the variance report Google Docs link, sent back to the original requester with one attachment — confirming the full end-to-end workflow executed successfully.

Project Details

This project automates the daily variance reporting process using an AI agentic workflow in n8n. When an email arrives with a CSV attachment containing new system data, the workflow triggers automatically and extracts the CSV content. An AI Agent powered by Google Gemini then queries the existing PostgreSQL database and performs a smart comparison across multiple dimensions — total rows, column names, numeric totals, and breakdowns at both overall and city level. Rather than using rigid logic-based rules, the AI agent handles the heavy lifting autonomously. The comparison results are written into a newly created Google Docs report, which is then shared via Google Drive. Finally, the report link is sent back to the original sender via Gmail. Key optimizations included using Aggregate and Edit Fields nodes to convert multi-value data into a single string for clean AI input. This project demonstrates skills in database integration, automated reporting and incremental workflow optimization

Project video

Automated Invoice Extraction & PDF Generation using n8n

Domain/Function: Data Engineering / Workflow Automation

Built an n8n workflow that automatically extracts invoice details from Gmail using AI, generates a professional PDF via APITemplate.io, saves a draft email with the attachment, and logs all invoice data to Google Sheets — zero manual effort.

View Project
Engagement

Automated Invoice Extraction & PDF Generation using n8n

Live n8n workflow: Gmail Trigger → Information Extractor (Gemini) → Create PDF → Create Gmail Draft → Append Row in Google Sheets. All nodes executed successfully with 1 item passing through each step.

Auto-generated PDF invoice showing customer name, course, amount, date, and invoice number — created via APITemplate.io from AI-extracted Gmail data, with payment details and grand total included.

Project Details

This project automates the end-to-end invoice handling process using n8n. Whenever a new invoice-related email arrives in Gmail, the workflow is triggered automatically. An AI Information Extractor node reads the email and pulls out key details such as customer name, course, amount, date, and invoice number. These details are then sent to APITemplate.io, which generates a professionally formatted PDF invoice using a pre-designed template. The PDF is attached to a draft email in Gmail, ready for review before sending — ensuring nothing goes out without approval. Finally, all invoice data is logged into a Google Sheet for record-keeping and tracking. The entire solution was designed with a step-by-step mindset, using ChatGPT as a second brain to plan each automation stage before building. This project demonstrates practical skills in workflow automation, API integration, AI-powered data extraction, and cloud-based document generation.

Project video

Machine Learning Project

Brazilian E-commerce Churn Prediction

Domain/Function: E-commerce / Retail Analytics

End-to-end churn pipeline on Databricks using Medallion Architecture. Engineered 25+ RFM and behavioral features. Compared Logistic Regression, Decision Tree, Random Forest with MLflow. Decision Tree achieved ROC-AUC 0.75, F1 0.39. Identified high-risk segments and simulated retention campaign with

View Project
Engagement

View Project
on GitHUB

ROC curves confirm strong ranking ability. Tuned thresholds boost churn recall, prioritising valuable churn capture over overall accuracy.

SQL dashboards provide: risk distribution across the full customer base top Brazilian states with the highest concentration of high-risk users churn patterns by lifecycle and value segment This supports targeted and region-aware marketing actions.

Project Details

This project implements an end-to-end churn prediction pipeline using the Databricks Lakehouse platform. Raw Brazilian e-commerce data is ingested into Bronze Delta tables, transformed into a Customer 360 view in Silver, and 25+ AI-ready features including RFM, delivery experience, review patterns, and geographic spending context are engineered in Gold. Multiple classifiers including Logistic Regression, Decision Tree, and Random Forest are trained and tracked via MLflow. The final Decision Tree model achieves ROC-AUC 0.75 and F1 0.39 on highly imbalanced churn data. Business insights are extracted using Databricks SQL, revealing that new low-spending customers are the highest churn-risk group. A targeted retention campaign is simulated, demonstrating 1.11× ROI with only 25% success rate. The project showcases full lifecycle data science on modern lakehouse architecture.

Project video

Deep Learning Project

Vehicle Damage Detection

Domain/Function: Computer Vision

Automated vehicle damage detection using ResNet50. Classifies Front/Rear damage severity (Normal/Breakage/Crushed) from single images. Built with TensorFlow, FastAPI & Streamlit. 80% validation accuracy. Reduces manual inspection bias & delays.

View Project
Engagement

View Project
on GitHUB

Streamlit app interface showing damage prediction for a car image

Project Details

Vehicle damage assessment remains manual, time-consuming, and inconsistent — directly impacting insurance claims, resale pricing, and repair workflows for companies like VROOM Cars.

I developed a deep learning proof-of-concept that automates damage classification through an interactive web app. Built on ResNet50 with transfer learning and trained on ~1,700 images of third-quarter front and rear vehicle views, it classifies images into six categories: Front/Rear Normal, Front/Rear Crushed, and Front/Rear Breakage, achieving ~80% validation accuracy.

Users simply upload a car image; the app instantly predicts damage location and severity. This reduces human bias, accelerates claims, and enables scalable assessment.

Tech Stack: TensorFlow, FastAPI, Streamlit. The modular system allows easy retraining and deployment, demonstrating how computer vision can transform automotive inspection.

Project video

Machine Learning Project

Credit Risk Prediction System

Domain/Function: Finance / Banking

Production-ready credit risk classification system. Implemented Logistic Regression, Random Forest, XGBoost. Evaluated using confusion matrix, ROC-AUC, KS statistic, decile analysis. Engineered credit utilization, delinquency ratio, average spend. Deployed Streamlit app for real-time prediction.

View Project
Engagement

View Project
on GitHUB

This helps explain which financial behaviors contribute most to default risk.

KS statistic measures model separation power between defaulters and non-defaulters Decile table shows concentration of defaulters in high-risk groups Higher KS values indicate strong discriminatory performance

Project Details

A production-ready machine learning system for credit risk prediction that classifies loan applicants as default or non-default. Multiple classification models including Logistic Regression, Random Forest, and XGBoost are implemented and compared. Model evaluation follows strict banking industry standards using confusion matrix, ROC-AUC, KS statistic, and decile analysis. Key financial features such as credit utilization ratio, delinquency ratio, and average spend are engineered to capture customer risk behavior. Feature importance analysis reveals that credit utilization and delinquency history are the strongest default predictors. The final model is deployed via an interactive Streamlit web application, enabling real-time credit risk scoring with explainable AI components including feature importance charts and performance metrics dashboard. This project demonstrates practical financial risk modeling for fintech and banking applications.

Project video

Machine Learning Project

Health Insurance Premium Prediction System

Domain/Function: Healthcare / Insurance

Dual-model regression system using age-based segmentation. Young group (≤25): Ridge Regression, R² 0.9887. Adult group (>25): XGBoost, R² 0.9935. Engineered 12 parameters including medical risk scoring. Streamlit app delivers real-time premium estimates.

View Project
Engagement

View Project
on GitHUB

Health Insurance Premium Prediction System

Adult Group Feature Impact Insurance plan, age, and medical risk dominate premium estimation for adults.

Young Group Feature Impact Insurance plan and genetic risk are the most influential factors for younger individuals.

Project Details

This project introduces a novel age-based dual-model regression strategy for health insurance premium prediction. Unlike traditional single-model approaches, this system recognizes that younger and adult populations exhibit fundamentally different risk patterns. For individuals aged 25 and under, Ridge Regression achieves R² 0.9887 with insurance plan and genetic risk as primary drivers. For adults over 25, XGBoost Regressor achieves R² 0.9935 with age, medical history, and insurance plan dominating. Twelve input parameters are engineered including an intelligent medical risk scoring system that calculates disease weights dynamically. Feature importance analysis provides full transparency into pricing factors. The production-ready Streamlit web application delivers instant premium estimates in real-time with tight error margins. This project demonstrates how domain-aware model design significantly outperforms generic regression approaches in healthcare pricing.

Project video

Python Project

Expense Tracking System

Domain/Function: Finance

An expense tracker with Streamlit frontend and FastAPI backend. SQLite database. Comprehensive Pytest test suite. Modular structure for scalability. Complete setup documentation.

View Project
Engagement

View Project
on GitHUB

Analysis_by_month

Analysis_by_Category

Project Details

An expense tracking system built with Streamlit frontend and FastAPI backend. The application implements RESTful API architecture for robust financial data management, enabling users to record, categorize, and analyze personal expenses. The backend server is built with FastAPI and Uvicorn, providing high-performance asynchronous request handling with SQLite database persistence. The Streamlit frontend delivers an intuitive user interface for expense entry, category management, and spending visualization. A comprehensive testing suite using Pytest achieves high code coverage across both frontend and backend components. The project follows production-grade software engineering practices including modular architecture (separate frontend/backend/test directories), dependency management via requirements.txt, and detailed setup documentation. This project demonstrates full-stack development capabilities essential for data science deployment and ML system integration.