Kigali Water Network Digital Twin (GIS)

Upanzi Network (Kigali, Rwanda) — July 2025 to Dec 2025
This project involved building a comprehensive digital twin for Kigali's water distribution network. The process began with structuring and cleaning large-scale GIS and utility datasets, ensuring that every node, pipe, and sensor was accurately mapped and validated. The digital twin enabled real-time monitoring, leak detection, and predictive maintenance, supporting both engineering decisions and field operations.

System Dashboard Overview:
The main dashboard (see below) provided a real-time overview of the entire water network, including node status, battery levels, flow rates, and system alerts. This interface empowered operators to quickly identify issues such as offline nodes or abnormal readings, and to drill down into individual node performance.

System Architecture:
The architecture of the cyber-physical system is illustrated below. It shows the flow of data from distributed sensor nodes through gateways to the cloud, where data is processed, visualized, and used for decision-making. This layered approach ensured robust data collection, secure transmission, and actionable insights for water quality and leak detection.

Project Impact:
By integrating GIS, IoT, and advanced analytics, the digital twin provided actionable intelligence for reducing non-revenue water, improving response times to leaks, and optimizing maintenance schedules. The system's modular design allows for future expansion, including integration with additional sensors and predictive analytics modules.

3D Prosthetic Reconstruction (AI/CV)

Carnegie Mellon University Africa (Kigali, Rwanda) — 2024
This project pioneered a workflow for reconstructing prosthetic models using AI and computer vision. The process started with smartphone-based data collection, followed by advanced preprocessing and segmentation to extract relevant features for 3D modeling.

Project Process Overview:
The diagram below illustrates the step-by-step process, from image acquisition to 3D printing. Each stage was optimized for accuracy and speed, ensuring high-quality prosthetic prototypes.

Reduced Time & Efficiency:
By leveraging AI-driven segmentation and reconstruction, the workflow significantly reduced the time required for prototyping. The chart below highlights the efficiency gains achieved compared to traditional methods.

Final 3D Model:
The resulting prosthetic model, shown below, was evaluated for fit and durability, and successfully prototyped using 3D printing. This approach enabled rapid iteration and customization for individual patient needs.

Impact & Publications:
The project contributed to peer-reviewed publications (IDETC-CIE ASME 2024; ASME Journal of Mechanical Design) and demonstrated the potential for scalable, patient-specific prosthetic solutions using AI and computer vision.

Taxi Fare Prediction (GCP + XGBoost)

Cloud Computing capstone (CMU-Africa) — Oct 2024 to Dec 2024
Delivered an end-to-end, reproducible pipeline for taxi-fare estimation using NYC taxi records and GCP-hosted services. The project focused on robust preprocessing, strong feature engineering, model tuning, and low-latency deployment for online predictions.

• Data & Preprocessing: Ingested and cleaned large-scale NYC taxi datasets (pickup/dropoff timestamps, coordinates, passenger counts, fares). Implemented validation checks, outlier filtering, and scalable feature pipelines to ensure reproducible training and evaluation.
• Feature Engineering: Built time-based (hour, day-of-week), spatial (H3/geohash bins, distance, bearing), and contextual features (passenger count, estimated traffic proxies). Engineered features to reduce bias and improve generalization across boroughs.
• Modeling: Trained and tuned an XGBoost model with cross-validation and hyperparameter search. Used careful validation splits to avoid temporal leakage and to measure realistic generalization performance.
• Deployment: Packaged the model and serving logic in a Flask microservice and deployed to Google App Engine for horizontally scalable, low-latency inference (production latency target < 200 ms).
• GCP Tools: Used AI Platform for managed training experiments, Cloud Storage for dataset and model artifacts, and App Engine for serving. Integrated logging and basic monitoring to track request rates and latencies.
• Extras: Added landmark recognition using Cloud Vision API to enhance contextual routing and optional AutoML prototypes for rapid experimentation and model comparisons.
• Reproducibility & Pipeline Integration: Automated preprocessing and model training steps so experiments can be re-run end-to-end; packaged artifacts with versioning to support retraining and audits.

Skills & tools: GCP (AI Platform, App Engine, Cloud Storage), XGBoost, Flask, Cloud Vision API, AutoML, Python, feature engineering, and pipeline integration.

Unified Data Aggregation and Visualization Pipeline

Unified Data Aggregation and Visualization Pipeline: Scalable Data Integration and Analytics Framework — Nov 2024 to Dec 2024
Carnegie Mellon University Africa, Kigali, Rwanda
Skills & Tools: Python, Prefect, Dash, Plotly, MongoDB, PostgreSQL, ETL, Big Data Analytics

• Developed an automated ETL pipeline using Prefect to extract, transform, and load data from diverse sources.
• Integrated data from PostgreSQL and MongoDB into a unified dataset, ensuring consistency across heterogeneous sources.
• Built an interactive dashboard with Dash and Plotly Express to visualize key metrics and generate actionable insights for stakeholders.
• Designed the framework for future scalability with the potential to incorporate advanced tools such as Kafka and Cassandra for real-time processing.

The pipeline emphasized data quality, reproducibility, and observability: schema normalization, deduplication, validation checks, and lightweight monitoring were implemented to ensure reliable downstream analytics.

Smart Sorter: Neural Network-Enhanced Recycling Bin

Smart Sorter: Neural Network-Enhanced Recycling Bin — Feb 2024 to Apr 2024
Carnegie Mellon University Africa, Kigali, Rwanda
Skills & Tools: Python, PyTorch, ResNet50, Stable Diffusion Model (SDM), Transfer Learning, Data Augmentation, Model Optimization, Image Classification, Dataset Engineering, Cross-Validation

• Final project for the Introduction to Deep Learning course, focused on robust, real-time waste classification for edge deployment.
• Developed a real-time waste classification system using a modified ResNet50 backbone optimized for inference on constrained devices.
• Enhanced dataset diversity using a Stable Diffusion Model (SDM): trained the SDM to generate realistic synthetic samples, applied diffusion and autoregressive techniques to diversify rare classes, and used these samples to improve model robustness.
• Applied transfer learning, targeted data augmentation, and model optimization (pruning/quantization experiments) to reduce latency while preserving accuracy for edge deployment.
• Achieved 97% classification accuracy after integrating synthetic data and transfer learning; validated results on the TrashNet dataset using cross-validation and standard metrics (precision, recall, F1).

Author: Jean Rodrigue Bitsinda Ikuzwe — Feb 16, 2026
The project emphasized dataset engineering, reproducible training pipelines, and practical edge considerations (latency, memory). Results include trained model artifacts and code to reproduce experiments.

Developing a Scalable Machine Learning-Based Potential Energy Surface (PES) Fit for Molecules: Sum-of-Product Approach

ICTP-EAIFR, Kigali, Rwanda

Developed Python software to construct machine-learned Potential Energy Surfaces (PES) in a sum-of-product representation for grid-based quantum-dynamical simulations, targeting spectroscopic-quality fits (RMSE ≈ 1 cm⁻¹) while maintaining scalability and symmetry awareness.

Approach

Investigated Atomic Cluster Expansion (ACE), Message-Passing Neural Networks, Kernel Ridge Regression, and Gaussian Process Regression; explored force-field-like functional forms and evaluated fits via quantum dynamics tests on Van der Waals, large-amplitude, and semi-rigid molecules.

Skills & Tools

Machine Learning, Python, TensorFlow, Model Optimization, Cluster Computing, Computational Chemistry, Neural Networks

Highlights

• Constructed ML-based SOP PES representations and compared NN/KRR/GPR approaches.
• Explored HDMR and symmetry-aware techniques to improve scalability.
• Validated fitted PES within quantum-dynamical simulations to assess spectroscopic accuracy.