Machine Learning

Fleet Driver Fatigue & Drowsiness Simulator

Architected and engineered an intelligent, computer vision-driven driver safety simulator designed to mitigate commercial fleet accidents by detecting operator fatigue, micro-sleep episodes, and distracted driving patterns in real-time. Deployed a high-throughput video processing pipeline that ingests live webcam frames to track facial topography. Utilizing MediaPipe Face Mesh alongside OpenCV, the system computes structural geometric metrics—such as Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR)—to detect prolonged eyelid closure, frequent yawning, and critical head-tilt deviations, triggering instantaneous multi-sensory alerts upon threshold violation.

Metrics at a Glance:

Key Architectural Achievements:

• Real-Time Facial Landmark Mesh: Implemented high-density facial landmark localization utilizing MediaPipe to track 468 discrete anatomical keypoints, guaranteeing consistent micro-expression monitoring under varying ambient in-cabin lighting conditions.
• Geometric Drowsiness Scoring: Engineered spatial coordinate algorithms to dynamically compute Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR), mapping chronological multi-frame thresholds to confidently differentiate standard blinking from hazardous micro-sleep states.
• State Machine Alert Logic: Developed a low-latency temporal state-tracking engine that tracks consecutive frame infractions, triggering cascading defensive actions from "Warning: Tiredness Detected" to "Critical Alert: Driver Sleeping" with audio-visual notifications.
• Asynchronous Thread Ingestion: Built an optimized frame ingestion pipeline using OpenCV, cleanly decoupling the camera data streaming layer from the deep learning feature-inference thread to completely eliminate visual stutter and UI freezing.
• Safety Telemetry Dashboard: Designed a clean, web-based simulator control layout providing telemetry monitoring, real-time threshold configuration sliders, operational log auditing, and immediate behavioral feedback loops.

AgriIntel Platform: End-to-End Architectural Data Flow & Predictive Intelligence Layer

Architected and engineered an intelligent, enterprise-grade agricultural data platform designed to optimize regional farming strategies and forecast commodity pricing using distributed machine learning. Built on a FastAPI Engine application layer, the system utilizes an asynchronous ingestion worker to orchestrate data pipelines across a decoupled infrastructure. The platform pairs a real-time geospatial processing layer with a dedicated predictive intelligence cluster to ingest telemetry data from edge-deployed localized farming units, aggregate geographic boundaries, and issue low-latency cellular agricultural advisories.

Metrics at a Glance:

Key Architectural Achievements:

• Predictive Intelligence Layer: Developed an analytics core deploying an Ensemble Crop Classifier (Random Forest) to evaluate agronomic yield potentials alongside a Time-Series Pricing Forecaster (Exponential Smoothing) for localized forward pricing.
• Geospatial Density Aggregation: Engineered a clustering system using DBSCAN Matrix Clustering to group disjointed farmlands and dynamically map spatial arrays across regional procurement hubs.
• Asynchronous Ingestion Worker: Built a non-blocking background task payload delivery system that decouples edge-device telemetry submissions from heavy machine learning model evaluation and matrix calculations.
• Resilient Data & Persistence Layer: Structured a multi-tier storage engine featuring an In-Memory Regional Geocode Cache (Redis/GeoJSON), an Write-Ahead Logging SQLite Shared Ledger, and an active Data Streaming Pipeline.
• Omnichannel Presentation Clients: Designed a hybrid client interface featuring a Web Management Dashboard (HTML5 / Leaflet.js Vector Canvas) for logistics planners alongside a Mobile Cellular Network gateway for low-bandwidth cellular advisories.

AI Cadastral, Multi-City Suite — Decentralized Predictive Intelligence Ecosystem

Architected and designed a highly robust, enterprise-grade decentralized platform engineered for multi-city analytical evaluation, predictive tracking, and immutable logging. Built upon a structured multi-layer architecture, the system coordinates secure user access through presentation dashboards, executes gateway logic via a core application layer, and powers deep model evaluation inside an isolated predictive analytics engine. By cleanly decoupling training, live inference, and validation pipelines, the system processes cross-city metric inputs while ensuring end-to-end data integrity through a centralized database paired with an immutable decentralized shared ledger.

Metrics at a Glance:

Key Architectural Achievements:

• Predictive Analytics & Intelligence Layer: Engineered a dedicated, multi-stage machine learning runtime environment using Scikit-Learn that completely decouples the training pipeline engine from the real-time inference and evaluation engines.
• Decentralized Shared Ledger: Integrated a secure, append-only immutable shared ledger directly beneath the standard database persistence layer to establish cryptographic accountability for all audit log records and evaluation metrics.
• Asynchronous Core Application Gateway: Designed a secure API Gateway and logic handler that filters client presentation requests, enforces dual-token security, and safely routes structured spatial data packages to the background ML execution pools.
• Isolated Model Validation Pipeline: Established a standalone evaluation engine within the analytics cluster to continuously compute model drift metrics against fresh database transactions before promoting models to active inference states.
• Secure Client Access Management: Programmed a tiered Presentation and Client layer featuring isolated operation portals and management dashboards to cleanly separate public analytical data consumption from deep system maintenance tasks.