AI-Powered HVAC Predictive Maintenance and Label Recognition Platform

The brief

Our client, an American startup, has developed an IoT HVAC solution that combines smart cleaning devices with web and mobile applications. The devices automatically clean A/C drain lines of bacteria and blockages, while the apps let B2C and B2B users easily install, monitor, and manage the devices, as well as order and replace cleaning cartridges.

As the product began expanding into the B2B market, our client noticed another opportunity. Professional HVAC integrators who install and maintain heating and cooling systems often face one common information gap. When arriving at a site, they rarely know what exact models of air handlers they will be dealing with, their warranty conditions, or service lifespans. This lack of information makes planning preventive maintenance more difficult, increases costs, and downtime.

To address this, our client proposed a new feature: let users and integrators take a photo of the air handler label during device setup. The platform would then recognize the model and extract key specifications, warranty information, and maintenance data. With this feature, predictive maintenance for HVAC systems would become significantly easier.

Challenge

Building the air handler brand identification feature was far from straightforward.

HVAC equipment comes from over 15 major manufacturers and many lesser-known ones, each with dozens of models. Therefore, we decided that creating a comprehensive reference database manually wasn’t sustainable. Instead, we chose AI as the only viable path forward.

The challenge was to design an air handler model and brand identification AI solution that could:

• Accept user-uploaded photos of HVAC air handler labels.
• Identify the air handler brand and model as a starting point.
• Extract a broader set of technical and service-related details, including:
  • Brand
  • Model number
  • Serial number
  • Product number
  • Manufacturing date
  • Tonnage
  • Electrical details
  • Horsepower
  • SEER rating
  • Warranty status and expiry date
  • Filter configuration and recommended size

Our goal for HVAC equipment data extraction from images was to get as much useful data as possible from every air handler photo. To do this while keeping costs down, we decided to develop an algorithm that captures all the data and significantly reduces analysis expenses.

Solution

To tackle the challenges and meet the set goals, we developed a multi-layered label recognition algorithm powered by AI. Here is how it works:

1. Photo submission. End users upload a photo of the air handler nameplate through the mobile or web app.
2. Task creation. The backend creates a task for photo description generation and places it in the Google Cloud Tasks Queue.
3. Asynchronous execution. Tasks are handled one by one, which ensures stable performance and optimized resource usage.
4. Initial analysis. The system uses a large language model (LLM) for the initial AI photo analysis for B2B HVAC integrators. LLM extracts such key details as brand, model number, serial number, and more.
5. Description return. The system returns the description with data points extracted from the photo.
6. Further analysis. The received description is analyzed to see if it contains the necessary data, primarily the brand and model.
   1. If the air handler model has already been processed, the system pulls details from the cache instead of re-analyzing the photo. This makes recognition faster and reduces costs for models previously identified and stored in the database.
   2. If it’s a new model, the system enriches the data using either Vertex AI grounding for HVAC data extraction (our curated library of manuals, PDFs, and photos) or Google Search grounding for additional validation.
7. Data clarification. The system returns a JSON file with all the data fields required, or at least the most critical ones.
8. Data normalization. The system cleans, standardizes, and validates all data before saving it in the database.
   1. Invalid or duplicate photos (e.g., of our own device) are filtered out.
   2. Manufacturer names are standardized.
   3. Manufacturing date is determined either from the label itself or by decoding the serial number with brand-specific rules (e.g., Trane’s serial code system).
9. Warranty status and estimated expiry date are calculated, even if the manufacturing date isn’t explicitly listed.
10. Data persistence. All structured HVAC metadata is stored in the database of our IoT-enabled HVAC monitoring platform.

Results

The AI-powered HVAC label recognition solution we delivered successfully transformed our client’s business.

• For B2B integrators, it provides immediate, actionable intelligence on HVAC units before arriving on-site.
• For end users, it offers clarity about their HVAC systems, whether it’s HVAC warranty status detection or recommended maintenance.
• For our client, it opens a new revenue stream, a competitive edge, and cost optimization.

In particular, our air handler model number recognition solution balances accuracy with cost efficiency by using a tiered strategy, with $0.000093–$0.000279 for cache retrieval, $0.002593–$0.007779 for Vertex AI grounding, and $0.035093-$0.105279 for Google Search grounding.