Nivara
Nivara: AI energy intelligence for Indian campuses that turns IoT telemetry into anomaly alerts, forecasts, and work orders, tested on simulated data.
- Problem, approach and result explained
- Public source code
- Expert reviewed· 6 reviews
5 years
Continuous Timeseries Historical Data
4 layer
Physical Cutoff for cost saving
96.3%
Forward Forecasting Accuracy Tested Against Historical Data
Overview
The problem
Campuses, hospitals, government offices and industrial estates in India lose a large share of their utility spend to waste nobody sees: baseload that never drops at night, HVAC running in empty rooms, underground water leaks, and demand and power factor penalties. Monitoring tools show charts, but nobody acts on them or checks that the savings were real. The real data needed to study this is private, so the problem is also how to build and test a decision system without it.
The approach
I built Nivara around one flow: problem, data, AI, insight, action, rather than data straight to a dashboard. I went deep on energy first. A simulator produces 28 days of 15-minute readings with injected faults, which pass through a FastAPI ingestion layer with quality checks, then into Ridge and Prophet forecasts and an anomaly engine (rules, statistical drift, Isolation Forest). Each anomaly gets a “Why?” explanation, a recommendation and a work order with an SLA timer. Savings are framed with IPMVP-style verification, and a separate integrated dashboard holds the other PS modules at lower resolution.
The result
The energy pipeline works end to end on synthetic data: ingestion, dual forecasting with what-if sliders, anomaly detection checked against injected ground truth, recommendations, work-order queue, one-click PDF report and a Gemini assistant. It runs on four facility datasets (university, hospital, government office, industrial campus) with 76/76 tests passing. The other modules (air quality, waste, water, traffic) are only partly built: they run on lower-resolution data, and water is a leak proxy rather than a full night-flow model. No real hardware is connected and no savings have been measured on a live site.
Reflection
Choosing one module and going deep beat spreading thin, and I’d make that choice again. The biggest limit is that everything rests on simulated data and 28 days of history, so the results show the system’s direction and design, not proven accuracy or real-world savings. If I continued, I’d connect real meters at a single campus to test it, build the other modules to the same depth, and replace any claimed savings with measured before-and-after results.
Submitted to Sustainable Facility & Estate Intelligence
- Your idea in one line
An autonomous 3D digital-twin and energy intelligence platform for university campus administrations that detects equipment faults in real-time, optimizes rooftop solar/battery microgrids, and automates sub-meter departmental billing to eliminate 20–35% of institutional resource waste.
- What will be working by the finale?
1. Live 3D Campus Digital Twin (MapboxGL): Interactive geospatial visualization of campus facilities with dense sub-meter pins (HVAC, lighting, pumps, computer labs) showing real-time power draw and pulsing red anomaly alerts for the latest date. 2. 5-Layer Autonomous Fault Detection & Diagnosis (FDD): Hybrid rule-based and unsupervised Machine Learning (Isolation Forest) pipeline that detects equipment left on, dry-running water pumps, and baseline deviations with confidence scores. 3. Automated Departmental Chargeback Ledger: Sub-meter billing engine that computes exact electricity costs and flags after-hours waste per academic department (CS, Mechanical, Admin, Hostels, etc.) instead of flat square-foot splits. 4. 24-Hour ML Load Forecaster & "What-If" Policy Simulator: Ridge-regression predictive model with confidence bands and interactive scenario sliders (heatwave temperature offset and evening AC shutdown curfews). 5. Solar & BESS Microgrid Peak-Shaving Simulator: Automated economic dispatch algorithm managing an 800 kW solar array and 2,000 kWh battery storage system with State-of-Charge (SOC) cycle health tracking. 6. Executive ESG Carbon Compliance Dashboard: Institutional dashboard calculating net carbon emissions (CEA grid factors) and downloadable one-click compliance audit reports. 7. NLP Complaint Correlation & Gemini AI Facility Copilot: Conversational assistant matching occupant complaints with simultaneous electrical faults and answering natural-language facility queries.
- Tech stack and hardware you plan to use
• Languages & Core Frameworks: Python 3.12, Streamlit, FastAPI, Pydantic, NumPy, Pandas. • Visualization & Digital Twin: Mapbox GL JS (3D terrain and polygon extrusion), Plotly (Express & Graph Objects), Tailwind CSS. • Machine Learning & AI: Scikit-Learn (IsolationForest, Ridge, StandardScaler), Google Gemini API • Database & Messaging Architecture: SQLite, REST APIs, AsyncIO Replay Engine (simulating live MQTT/Kafka telemetry streams). • Hardware & IoT Sensors (Planned / Simulated in POC): - Energy Meters: Smart multi-function energy sub-meters (Schneider / Modbus RS-485 meters). - Gateways & Boards: Raspberry Pi 4 / ESP32 industrial IoT gateways running Modbus-to-MQTT bridges. - Sensors: Split-core Current Transformers (CTs), digital water flow meters (pulse/hall effect for pumps), and ambient DHT22/DS18B20 temperature/humidity sensors.
- What have you built or tested so far?
We have built and tested a working end-to-end prototype using Python 3.12 and Streamlit. 1. Synthetic Data Simulator: Generated 290,304 five-minute telemetry records over 28 days for 36 assets across six university buildings. Simulates occupancy, weather, power consumption, power factor, pump flow and five operational faults, including after-hours usage, pump failures, sensor dropouts and duplicate packets. 2. Five-Layer Anomaly Detection: Implemented Ridge-regression baselines, occupancy-based rules, physical equipment checks, sensor-health monitoring and Scikit-learn Isolation Forest. Includes four-hour alert deduplication and confidence scoring. 3. Forecasting & Simulation: Built a 24-hour Ridge-regression forecaster using a rolling 21-day window, rolling statistics and academic-calendar context. Added interactive temperature-offset and policy-simulation controls.
- Pitch deck (PDF)
- Pitch deck (PDF)PDF
- GitHub repository
- github.com/paramjeetn/BrainChain-BPUT
- What's working now?
Working: the energy pipeline runs end to end on synthetic data. It covers 28 days of fault-injected 15-minute telemetry, FastAPI ingestion with quality checks, Ridge + Prophet forecasts with what-if sliders, anomaly detection with “Why?” explanations checked against injected ground truth, recommendations, a work-order queue with SLA timers, a PDF report, and a Gemini assistant. It runs on 4 facility datasets with 76/76 tests passing. Partly working: the air quality, waste, water and traffic modules exist at lower resolution and are not taken deep. Water is a leak proxy, and scenario simulation is strong for energy only. Not done: no real hardware is connected, and no savings have been measured on a live site.
- GitHub repository
- github.com/paramjeetn/nivara
- Deck or report (PDF)
- Deck or report (PDF)PDF
AI tools used
Claude helped with architecture planning, the pitch script, and reviewing the project for gaps and overclaims. Gemini powers the in-app assistant that returns structured actions from the dashboard data. All code, data design, and testing decisions were mine.