Skip to content
Back to Projects

AI-Powered BACnet Datapoint Normalization

Days to hours: AI-driven site onboarding for smart buildings

2025–2026 Energy / Smart Buildings / AI
See all 4 screens

Top outcome

Site onboarding cut from days to hours, with roughly 80% lower cloud token costs.

Stack

Ollama · Claude API · Python · SQLite

Overview

An AI-powered system that transforms chaotic BACnet datapoint naming into standardized formats across BRICK, Project Haystack, and W3C Web-of-Things (WoT) schemas. The engine uses primarily Ollama locally for the heavy lifting, supplemented by Claude API for complex edge cases, with domain-specific prompt templates to understand building semantics - equipment types, floor plans, zones, and measurement types - and automatically normalize thousands of datapoints per building. A human validation loop ensures production accuracy while reducing site onboarding from days to hours, unlocking downstream analytics and automation.

How it works

  1. Every building names its equipment points its own way.
  2. A local model proposes a standard name. Claude takes the hard cases.
  3. A person checks the result before it is used.
  4. Clean BRICK, Haystack and WoT names: onboarding in hours, not days.

The Problem

Commercial buildings have hundreds of heterogeneous BACnet devices with wildly varying naming conventions, making cross-building data analysis impossible.

The Solution

AI system that learns building semantics and auto-normalizes datapoint names to BRICK, Project Haystack, and W3C Web-of-Things (WoT) standards with a human validation loop.

Impact

  • 1Site onboarding reduced from 3–5 days to 2–4 hours
  • 2Enabled automated energy analytics across multi-building portfolios
  • 3Unlocked downstream analytics, automation, and cross-site benchmarking by producing clean, standardized metadata
Key Decisions
  • Built custom prompt templates for BACnet semantics (equipment type, floor, zone, measurement)
  • Implemented local AI via Ollama for the heavy lifting, significantly reducing cost while reserving Claude API for complex edge cases
  • Implemented human validation loop to catch edge cases (critical for production)
  • Used context windows efficiently to process building metadata
Lessons Learned
  • Domain-specific AI requires deep understanding of the domain (BACnet conventions, building operations)

  • Production AI systems need human oversight; don't automate the human out of the loop

  • Accurate metadata (building schematics, equipment names) is foundational

Screens

1 / 4