System online

ARCHITECTING
INTELLIGENCE

Building intelligent systems, edge devices, and data-driven solutions for the real world.

// Tech stack

Languages & ML

Python TensorFlow Scikit-Learn MATLAB Numpy Pandas

Web engineering

  • Streamlit
  • Flask / REST APIs
  • HTML / CSS
  • Hugging Face

Cloud & DevOps

Google Cloud Run Docker Git/GitHub

Hardware / IoT

Arduino, ESP32, Edge Computing & Embedded System

// Deployed systems

Full-Stack Development, AI Agents, and SaaS Prototypes.

MULTI-AGENT AI

Municipal Infrastructure Sentinel

City-scale triage using Multi-Agent AI to validate infrastructure reports. Features Gemini 2.5 Flash vision verification and geospatial prioritization logic.

PYTHON GEMINI FOLIUM
VOICE AI

Naija Legal Aid

Voice-First AI bridging the gap to complex law. Speaks Pidgin English, generates formal .docx demands instantly without cloud storage.

GTTS VOICE-ENGINE DOCX
FINTECH

Naija MarketSense

Multimodal AI agent for SMEs visualizing price spreads (arbitrage) on heatmaps. Interprets heavy Pidgin accents via "Wise Trader" persona.

PLOTLY MULTIMODAL GEMINI
RL AGENT

Autonomous Energy Grid

Deep Reinforcement Learning (PPO) agent stabilizing renewable microgrids. Achieved 5.5% cost savings vs standard controllers via optimal battery discharge strategies.

PYTORCH GYMNASIUM STABLE-BASELINES3

WiFi Performance Auditor

Full-stack bandwidth monitor using event-driven architecture to visualize latency trends and SLA violations in real-time.

Credential Leak Auditor

Security tool auditing employee emails against breach databases. Features a mock API layer to simulate "Have I Been Pwned" responses.

// Data lab

Medical Diagnosis, Interpretability (XAI), and Predictive Analytics.

ACC: 98.2%

Breast Cancer Diagnosis

Explainable AI (SHAP) to visualize why the model made a diagnosis.

LAUNCH APP
41 CLASSES

Disease Prediction

Multi-class prediction system supporting bulk CSV uploads for clinics.

LAUNCH APP
STACKING

Heart Disease Risk

Combining Logistic Regression and XGBoost for robust clinical models.

LAUNCH APP
AGRITECH

Crop Recommendation

Optimizing yield by analyzing soil N-P-K values for Nigerian farmers.

LAUNCH APP
F1: 0.99

NLP Spam Engine

Real-time SMS classification using Random Forest, TF-IDF, and SMOTE.

LAUNCH APP

// Publications & research

Peer-reviewed literature, pre-prints, and academic contributions hosted on Zenodo.

JOURNAL • ZENODO DOI: 10.5281/zenodo.18794765

Physics-Informed Domain Adaptation: A Digital Twin Approach to Fault Diagnosis in Data-Scarce Industrial Environments

Presents a Physics-Informed Domain Adaptation (PIDA) approach combining a high-precision Digital Twin model and a 1D-CNN. Achieved 100% precision and recall for Sim-to-Real fault diagnosis on the CWRU-bearing dataset in a heavily data-scarce setting.

READ PUBLICATION
JOURNAL • ZENODO DOI: 10.5281/zenodo.18794753

Binary Saturation Dynamics in Recurrent Spiking Neural Networks: Limitations of Homeostatic Regularization

Investigates Homeostatic L2 Regularization constraints on rate-coded RSNNs. Reveals that while networks achieve 100% classification accuracy, they settle into power-hungry, high-activity regimes, highlighting the critical need for hard neuromorphic sparsity constraints.

READ PUBLICATION
JOURNAL • ZENODO DOI: 10.5281/zenodo.18794732

Autonomous Active Vibration Control: A Deep Reinforcement Learning Approach for Dynamic PID Tuning in Non-Linear Systems

Proposes a self-sustaining active vibration control system using a Soft Actor-Critic DRL algorithm to dynamically adjust PID gains. Eradicates overshooting, reduces settling time by 40%, and optimizes energy efficiency by 17% compared to standard Ziegler-Nichols logic.

READ PUBLICATION
JOURNAL • ZENODO DOI: 10.5281/zenodo.18745734

Zero-Shot Industrial Fault Diagnosis via Physics-Informed Generative Adversarial Networks (PI-GANs): A Cross-Domain Validation Study

Tailors a PI-GAN for zero-shot predictive maintenance by integrating Ohm's and Newton's laws directly into the generator loss function. Delivered a zero-shot accuracy of 98.72% on unseen Inner Race faults and interdomain accuracy of 77.22% on Outer Race faults.

READ PUBLICATION
JOURNAL • ZENODO DOI: 10.5281/zenodo.18714046

Riemannian Transfer Learning for Cross-Subject BCI: A ShallowConvNet Approach to Zero-Calibration Motor Imagery Decoding

Addresses the calibration bottleneck in Brain-Computer Interfaces (BCIs) using Euclidean Alignment and SimCLR pre-training. The framework achieved 79.3% classification accuracy on unseen subjects with an ultra-low inference latency of 0.27ms.

READ PUBLICATION

> Who am I

I am an engineer with a passion for solving African-context problems using global technology. Optimizing Models and building systems that work in the real world.

$ CURRENT_TASK: ESP32 TinyML Optimization

Initiate communication

Fharuk147@gmail.com
LINKEDIN GITHUB | +234 903 043 8653

© 2026 Fharuk. System Status: Normal.