Built for HACKaMINeD 2026
Solar Intel is a full-stack intelligent solar plant management platform that combines:
- 🤖 ML Predictive Maintenance — 95.3% accuracy, AUC-ROC 0.981 (XGBoost)
- 💬 RAG Conversational AI — entity-aware Q&A grounded in live MongoDB fleet data
- 🧠 Agentic AI Workflows — 4-step autonomous pipeline: Data → Risk → Tickets → Summary
- 📊 GenAI Narratives — Groq Llama 3.3 70B generates plain-English operational reports
- 🌐 15-Language i18n — full real-time translation across every page
- 🐳 Production Ready — Docker multi-container, 20 unit tests, Vercel deployed
It doesn't just show data — it predicts failures before they happen, explains why in plain English, and auto-generates maintenance tickets.
| Pillar | Implementation |
|---|---|
| 🔍 Data Mining | Real solar telemetry ingested from CSV/Excel, 51-feature ML pipeline |
| 🤖 GenAI | Groq Llama 3.3 70B generates operational narratives from ML predictions |
| 🧠 RAG | Entity extraction → MongoDB context retrieval → grounded LLM Q&A with source citations |
| 🤖 Agentic AI | 4-step autonomous: Data Retrieval → Risk Assessment → Ticket Drafting → Executive Summary |
| 📊 ML Model | HistGradientBoosting classifier, SHAP explainability, FastAPI microservice |
| 🌍 Accessibility | 15-language real-time translation across every page |
| 🐳 Production Ready | Docker multi-container, 20 unit tests, Vercel + Docker deploy |
┌─────────────────────────────────────────────────────────────────┐
│ Frontend — Next.js 14 (App Router · React 18 · TailwindCSS) │
│ Dashboard · Plants · Inverters · Chat · AI Insights · +7 more │
├─────────────────────────────────────────────────────────────────┤
│ API Layer — 17 Next.js Route Handlers │
│ /api/dashboard /api/plants /api/chat /api/predict +13 more │
├─────────────────────────────────────────────────────────────────┤
│ Backend Services (TypeScript) │
│ Dashboard · Plant · Inverter · RAG · Agent · AI Advisor · ML │
├──────────────┬──────────────────┬───────────────┬──────────────┤
│ FastAPI ML │ MongoDB Atlas │ Groq Llama │ Open-Meteo │
│ :8000 │ (4 collections) │ 3.3 70B │ Weather API │
└──────────────┴──────────────────┴───────────────┴──────────────┘
- 3 solar plants, each with dedicated inverter fleets
- Plant-level aggregated metrics (total power, health score, risk level)
- Drill-down from plant → individual inverter telemetry
- Drag-and-drop file upload
- Auto-detects whether file contains plant or inverter data
- Column normalization handles camelCase, spaces, special characters
- Bulk upsert directly to MongoDB
- Model: XGBoost / HistGradientBoosting (95.3% accuracy · AUC-ROC 0.981)
- 29 features: raw telemetry + lag (1/4/8 step) + rolling stats + engineered ratios
- SHAP explanations — top-5 contributing factors per prediction
- FastAPI microservice —
/predict,/predict/batch,/predict/fleet
- RAG Mode: entity extraction → MongoDB context → grounded LLM response with source citations
- Agent Mode: autonomous 4-step workflow
- 📊 Data Retrieval — pulls live plant/inverter/telemetry context
⚠️ Risk Assessment — ML-powered risk evaluation for all inverters- 🎫 Ticket Drafting — auto-generates prioritized maintenance tickets
- 📋 Executive Summary — synthesized stakeholder-ready report
- Groq Llama 3.3 70B generates plain-English operational reports
- 3-tier fallback: ML + Groq → ML-only → rule-based (never fails)
- Real-time fleet KPIs with animated dashboards
- 48-hour solar generation forecast (weather-aware via Open-Meteo)
- Daily CO₂ avoidance tracking with real-world equivalents
en hi es fr de ja zh ar pt ru ko it nl tr pl
| Layer | Technology |
|---|---|
| Frontend | Next.js 14 (App Router), React 18, TailwindCSS, Framer Motion, Tremor |
| Backend | Next.js API Routes, Mongoose 9 |
| Database | MongoDB Atlas |
| ML Model | XGBoost + SHAP |
| ML API | FastAPI + Uvicorn (Python 3.12) |
| LLM | Groq — Llama 3.3 70B Versatile |
| Auth | NextAuth.js v4 (Google OAuth + Credentials) |
| i18n | 15 languages (static + Google Translate) |
| Import | PapaParse (CSV) + SheetJS (Excel) |
| Testing | Jest + ts-jest (20 tests, 3 suites) |
| Deployment | Vercel (web) · Docker multi-stage (containers) |
solar-intel/
├── src/
│ ├── app/ # Next.js pages (12 routes)
│ │ ├── page.tsx # Dashboard
│ │ ├── plants/ # Plant overview + CSV import
│ │ ├── chat/ # RAG + Agent chat UI
│ │ ├── ai-insights/ # GenAI health narratives
│ │ ├── analytics/ # Energy analytics
│ │ ├── anomalies/ # Anomaly detection
│ │ ├── forecast/ # 48-hr solar forecast
│ │ ├── carbon/ # CO₂ tracking
│ │ ├── maintenance/ # AI-prioritized task queue
│ │ ├── security/ # Grid sync quality
│ │ ├── settings/ # User settings
│ │ └── api/ # 17 Route Handlers
│ │ ├── chat/ # RAG + Agent endpoint
│ │ ├── predict/ # ML prediction proxy
│ │ ├── ai-advisor/ # GenAI narrative
│ │ ├── import/ # CSV/Excel import
│ │ ├── plants/ # Plant CRUD
│ │ └── ... # + 12 more
│ │
│ ├── backend/
│ │ ├── models/ # Mongoose schemas
│ │ │ ├── Plant.ts
│ │ │ ├── Inverter.ts
│ │ │ ├── Telemetry.ts
│ │ │ └── User.ts
│ │ └── services/ # Business logic (18 services)
│ │ ├── rag.service.ts # RAG pipeline
│ │ ├── agent.service.ts # Agentic 4-step workflow
│ │ ├── ml-prediction.service.ts
│ │ ├── ai-advisor.service.ts
│ │ └── ...
│ │
│ ├── components/ # React UI components
│ ├── lib/ # API client, auth, i18n utils
│ ├── types/ # TypeScript interfaces
│ └── __tests__/ # Jest unit tests (3 suites · 20 tests)
│
├── ml-service/ # FastAPI ML microservice
│ ├── main.py # /predict /predict/batch /predict/fleet
│ ├── predict.py # Full preprocessing + XGBoost inference
│ ├── solar_inverter_xgboost_model.pkl
│ └── requirements.txt
│
├── docs/
│ ├── ARCHITECTURE.md # Mermaid system diagrams + data flows
│ ├── MANUAL.md # Installation & execution manual
│ ├── PROMPT_DESIGN.md # Prompt iteration history
│ └── REPORT.md # Technical report (model rationale, GenAI choices)
│
├── scripts/seed-db.ts # Seeds 3 plants + 8 inverters + 3120 telemetry
├── Dockerfile # Multi-stage Next.js build
├── docker-compose.yml # Web + ML service
└── package.json
- Node.js 20+ · npm 9+
- Python 3.11+
- MongoDB Atlas account (free)
- Groq API key (free at console.groq.com)
git clone https://github.com/Jaimin2687/Solar-Intel.git
cd Solar-Intel
npm installCreate .env.local in the project root:
MONGODB_URI=mongodb+srv://<user>:<password>@<cluster>.mongodb.net/solar-intel
NEXTAUTH_SECRET=your_nextauth_secret
NEXTAUTH_URL=http://localhost:3000
GOOGLE_CLIENT_ID=your_google_client_id
GOOGLE_CLIENT_SECRET=your_google_client_secret
ML_SERVICE_URL=http://localhost:8000/predict
ML_TIMEOUT_MS=10000
GROQ_API_KEY=your_groq_api_key
GROQ_MODEL=llama-3.3-70b-versatile
EMAIL_FROM=your@gmail.com
EMAIL_APP_PASSWORD=your_gmail_app_passwordnpm run seed
# → Seeds 3 plants + 8 inverters + 3,120 telemetry recordscd ml-service
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 8000npm run dev
# → http://localhost:3000npm test
# → 3 suites · 20 tests · all passing ✅docker-compose up --build
# Web App → http://localhost:3000
# ML API → http://localhost:8000| Route | Method | Description |
|---|---|---|
/api/dashboard |
GET | Fleet KPIs, risk matrix, trend charts |
/api/plants |
GET | All plants with aggregated metrics |
/api/inverters |
GET/POST/PUT/DELETE | Inverter CRUD |
/api/telemetry |
GET/POST | Telemetry data |
/api/chat |
POST | RAG + Agent chat |
/api/predict |
POST | ML prediction proxy → FastAPI |
/api/ai-advisor |
GET | GenAI narrative (3-tier fallback) |
/api/import |
POST | CSV/Excel → MongoDB |
/api/analytics |
GET | Energy production analytics |
/api/anomalies |
GET | Deviation alerts |
/api/live-energy |
GET | Real-time energy feed |
/api/weather |
GET | 48-hr Open-Meteo forecast |
/api/grid |
GET | Grid sync metrics |
/api/translate |
POST | Batch i18n translation |
/api/email |
POST | Email notifications |
/api/health |
GET | System health check |
/api/auth/[...nextauth] |
* | Authentication |
| Suite | Tests | Coverage |
|---|---|---|
| ML Prediction | 4 | Feature mapping, defaults, normalization, 51-feature output |
| RAG Entity Extraction | 7 | Plant/inverter ID, keywords, case-insensitive, deduplication |
| Import Normalization | 9 | camelCase→snake_case, column detection, type inference |
| Total | 20 | All passing ✅ |
Task: Binary classification — healthy vs faulty inverter
Model: XGBoost classifier — 95.3% accuracy, AUC-ROC 0.981
Preprocessing Pipeline (inference replicates training exactly):
- Log1p transform on skewed features (PV1 current, PV2 power, lifetime kWh)
- Per-inverter z-score normalization — prevents inter-inverter bias
- Lag features — power & temp at 15min, 1hr, 2hr ago
- Rolling stats — 1hr & 4hr window mean + std
- Diff features — rate of change for power & temperature
- Engineered ratios — PV1/PV2 string balance, power/temp efficiency
Risk thresholds: LOW (0–0.3) · MEDIUM (0.3–0.6) · HIGH (0.6–0.85) · CRITICAL (0.85–1.0)
User Query → Entity Extraction (PLANT-xxx / INV-xxx / keywords)
→ MongoDB Context Retrieval (plants + inverters + telemetry + ML predictions)
→ Grounded Prompt Construction (with hallucination guardrails)
→ Groq Llama 3.3 70B → Response with source citations
Step 1: toolRetrieveData() → Pulls full fleet context from MongoDB
Step 2: toolRunRiskAssessment() → ML predictions for all inverters
Step 3: toolDraftTickets() → Creates maintenance tickets (risk ≥ 0.5)
Step 4: generateNarrative() → Groq synthesizes executive summary
Tier 1: ML predictions + Groq Llama 3.3 70B → Rich narrative
Tier 2: ML predictions only → Structured rule-based report
Tier 3: No ML data → Basic rule-based summary
npm run dev # Development server (HMR)
npm run build # Production build
npm start # Production server
npm run lint # ESLint
npm test # Jest (20 tests)
npm run seed # Seed database| Document | Description |
|---|---|
| ARCHITECTURE.md | Mermaid diagrams, data flows, component table |
| MANUAL.md | Full installation, dependency & execution guide |
| PROMPT_DESIGN.md | Prompt iteration history (before/after) |
| REPORT.md | Model rationale, GenAI choices, limitations |
| Page | Route | Description |
|---|---|---|
| Dashboard | / |
Fleet KPIs, risk matrix, trend charts |
| Plants | /plants |
Plant overview + CSV/Excel import |
| Chat | /chat |
RAG + Agent conversational AI |
| Inverters | /inverters |
Inverter fleet monitoring |
| AI Insights | /ai-insights |
GenAI health narratives |
| Analytics | /analytics |
Energy production analytics |
| Anomalies | /anomalies |
Deviation alerts & log |
| Forecast | /forecast |
48-hour solar forecast |
| Carbon | /carbon |
CO₂ avoidance tracking |
| Maintenance | /maintenance |
AI-prioritized task queue |
| Security | /security |
Grid sync quality |
| Settings | /settings |
Profile and preferences |
Current Limitations:
- RAG uses keyword matching (not vector similarity) — works well for structured IDs/metrics
- ML model trained on 3 plants — may need retraining for different solar technologies
- Agent steps are sequential — parallel tool execution would reduce latency
- No real-time SCADA streaming (batch per request, not WebSocket)
Future Improvements:
- Vector DB (Pinecone/Weaviate) for semantic RAG search
- Multi-model ensemble (HGBC + LSTM for time-series anomaly detection)
- Real-time SCADA integration via MQTT/Modbus
- CI/CD with automated model retraining on new data
- Multi-tenant architecture with role-based access per plant operator
Neev Modh — Built with ❤️ for HACKaMINeD 2026 🏆
☀️ Solar Intel — Predict failures before they happen. One watt at a time.