Hau Hu

Summary Meeting discussions evaluated AI platform viability for extracting construction data via contractor workflows and manufacturer integration. Platform Capabilities and Targets The platform automates bill of materials generation for electrical contractors by extracting symbols and schedules from construction drawings. Development efforts currently prioritize integrating single-line diagrams with project specifications. Technical Challenges and Strategy System accuracy relies on constraining Large Language Model inputs to specific document sections to mitigate high data variance. Integrating disparate specification formats remains a significant hurdle for comprehensive output. Market Fit and Applications Commercial contractors gain significant value from automated estimating and material planning, whereas industrial projects require strict brand adherence. The quotation team will likely utilize this tool for internal verification. Next steps [Alexei Schiopu] Demo System Progress: Send Hu Hao video demonstrating single line diagrams to specifications connection. Seek feedback regarding extracted data validity and accuracy. Details Introductions and Background: Alexei Schiopu and Hu Hao introduced themselves, discussing their past experiences and geographic connections to the Toronto area (00:00:00). Alexei Schiopu shared their professional background, which includes working as an electrician, studying electrical engineering, building software for 10 years at a company acquired by Autodesk, and transitioning into AI research for construction problems. Their current focus is on extracting structured data from construction drawings for estimating and other use cases like Quality Assurance/Quality Control (QAQC) and project management (00:03:13). Evolution of the AI Platform: Alexei Schiopu's platform initially focused on electrical elements, such as lighting, power, and communications, and is slowly expanding to include mechanical and plumbing facets of a commercial building, with current work on single-line diagrams. The platform is periodically opened to users for free two-week trials to gather feedback and value assessment, with the next opening scheduled for mid-May. The primary current use cases identified are for electrical contractors, manufacturers, and distributors, who need to generate bills of materials from blueprints (00:04:11). Hu Hao's Professional Background: Hu Hao, who also has an engineering background, shared their experience working with multiple manufacturers, including Eaton and Schneider. Their roles have ranged from design engineer (mechanical and electrical) to quotation, sales, and now application engineer lead, providing technical support for new projects across Canada (00:05:17). Initial Assessment of the Platform's Value Proposition: Alexei Schiopu sought feedback on whether converting single-line diagrams to a bill of materials makes sense and where its usefulness lies (00:05:17). Hu Hao noted that the method of building products differs across manufacturers, introducing complexity and variety, and inquired whether the system targets contractors or others (00:06:32). Target Users and Application Scenarios: Alexei Schiopu aims to target multiple user groups, starting with electrical contractors who need to count items like lights, power sources, and switchgear to generate a bill of materials (00:06:32). The system also applies to distributors and manufacturers, where sales personnel use blueprints to create bills of materials, particularly for specialized components like switchgear or panels. Alexei Schiopu is also considering making the technology available as an Application Programming Interface (API) for manufacturers who may be building similar tools internally, allowing them to ingest documents, hit the API for a bill of materials, and then handle internal processes like pricing (00:07:21). Challenges with Converting Single-Line Diagrams and Specifications: Hu Hao raised concerns about the technical difficulty, noting that single-line diagrams often fail to reflect the complexity of device requirements. They suggested the tool might be better suited for simplified products, such as lighting and small panels, but complex equipment is a challenge because specifications detailing enclosures and other requirements are often separate from the single line (00:08:24). Alexei Schiopu acknowledged the need to integrate specifications (using formats like CSI Division 2200 and 2600) but noted that project specifications are not standardized and vary depending on the consulting company (00:09:31) (00:13:31). Demonstration of the Data Extraction Platform: Alexei Schiopu demonstrated the current state of the platform, showing that uploaded documents take 15 to 30 minutes to process and provide an overview summary geared toward electrical contractors. The system extracts information by stitching drawings together, associating symbols (like lighting) with data found in schedules, including the tag, panel, and circuit, and can also calculate conduit lengths based on panel location (00:10:31). Alexei Schiopu is currently developing the single-line detection model, which will detect objects, build a hierarchical tree, and combine this data with specifications to create a comprehensive bill of materials (00:11:28). Handling High Variance in Project Data: Alexei Schiopu addressed concerns about the non-standard nature of project documentation, stating that every piece of extracted data, including panel schedules and drawing symbols, exhibits high variance (00:15:28). The strategy to achieve high accuracy using AI and Large Language Models involves constraining the input data by breaking up specifications into sections and asking direct questions, as giving the AI the entire specification package often leads to inaccuracies (00:16:23). Additionally, information gathered from the sheets (like panel names and types) will be used to improve the accuracy of questions asked against the constrained documentation (00:17:20). User Verification and Usefulness for Hu Hao's Role: Alexei Schiopu recognized that current users require a review state because trust in AI extraction is not yet universal, which led to the development of a user interface where everything can be verified (00:17:20). Hu Hao clarified that the tool would likely not be useful for their day-to-day work, which focuses on sales and helping engineering firms write specifications that meet client needs and comply with codes (value engineering). They suggested the tool would be valuable for the quotation team (00:18:48). Discussion of Value Engineering and Market Dynamics: Alexei Schiopu proposed that the tool could assist in value engineering by comparing different manufacturer models (e.g., Eaton vs. Schneider) that satisfy the design requirements at different price points. Hu Hao noted that large consulting companies already use internal software to generate customized specs and straightforward single-line diagrams (00:21:08). Furthermore, in commercial projects, contractors are not strictly required to follow the brand specified in the documents, allowing for value engineering, unlike in industrial projects (data centers, oil and gas, utilities) where end users are powerful and enforce strict adherence to a single brand for maintenance standardization (00:23:36). Project Value and Future Commitment: Hu Hao concluded that Alexei Schiopu's tool does not offer much value for highly direct industrial projects but confirmed its significant value for commercial project contractors seeking quick estimates, bill of materials creation, and labor hour planning (00:25:24). Alexei Schiopu committed to reaching out to Hu Hao once the single-line to specification integration is complete to receive feedback and verification (00:26:16) (00:27:59). Hu Hao agreed that their quotation team might use the tool and that the consulting engineer team would likely not, though the tool could potentially assist in QAQC checks for engineering mistakes (00:27:03).