ThapeloRanneileng/AI-Enhanced-CDMS-Web

AI-Enhanced CDMS for Lesotho — 4-model ML ensemble, Groq LLM explanations, QC Review Workspace, AES-256 security, Kubernetes. B.Eng Final Year Project, NUL 2026. 71/71 tests passing. AI-Enhanced Climate Data Management System for Lesotho > **B.Eng Computer Systems and Networks — Final Year Project** > > National University of Lesotho, Department of Mathematics and Computer Science > **Authors:** Thapelo Ranneileng | Motlatsi Masilo > **Target Organisation:** Lesotho Meteorological Service > **Base System:** Extended from Climsoft Web (open-source CDMS) --- ## What We Built This repository contains a **complete AI-enhanced extension** of Climsoft Web. We did not simply install Climsoft — we engineered the following systems on top of it. --- ## 1. LMS AI Pipeline (Python) **Location:** `front-end/pwa/aws-ingestion-layer/` - Trained a **4-model ensemble** on 893,398 historical LMS climate observations - Models: Z-Score baseline, Isolation Forest, One-Class SVM, Autoencoder (TensorFlow/Keras) - Ensemble decision layer: NORMAL / SUSPECT / FAILED classifications - **Groq API** (llama-3.3-70b-versatile) generates contextual reviewer explanations - Output: 714,816 scored predictions, 116,443 QC review handoff rows - Ensemble anomaly rate: 6.55% (11,700 review candidates from 178,704 scored) --- ## 2. QC Review Workspace (Angular — built from scratch) **Location:** `front-end/pwa/src/app/quality-control/` - Unified QC Review Queue: rule-based QC + ML anomaly decisions in one view - AI evidence panel: score, confidence, model agreement, Z-score signals, Groq explanation - **Approve / Override / Escalate** reviewer decision controls - Reviewer feedback loop: decisions persisted to PostgreSQL for Random Forest training - Demonstrated with LESBER07 tmax FAILED case: AI Score 0.947, confidence 96%, rolling Z-score -5.86 --- ## 3. AI Anomaly Management Centre (Angular — built from scratch) **Location:** `front-end/pwa/src/app/quality-control/ai-anomaly-management/` - Live dashboard: 893,398 clean rows, 714,816 predictions, Groq status - Model performance table (4 individual models + ensemble) - Groq-powered AI Climate QC Assistant with natural lan …