Artificial intelligence now writes software, interprets images, and appears in products people use every day. But as it takes on more real work, capability is not the same as reliability. Modern AI can still fabricate facts, misread context or cite nonexistent evidence, failures commonly known as hallucinations. Across building design and construction, those failures could contribute to a missed requirement, costly construction change, or safety risk.
In healthcare, they could influence a diagnosis or treatment decision with consequences for a patient’s life. Because professionals remain responsible for the outcome, the larger blast radius of a mistake makes them more cautious about adopting the technology.
At ArchiBoost AI in San Francisco, Malaysian founding engineer Yi Sien “Ian” Ku is helping the team confront that trust gap across building design and construction, where the consequences of a faulty conclusion can extend beyond a screen and into a physical building.
Built for Design and Construction Professionals
Ku joined ArchiBoost as its first technical hire to help tackle a problem that long predates AI: the slow and fragmented work of reviewing building designs. Before a project reaches construction, design and construction professionals must reconcile information across drawings, specifications and regulations to determine whether the design is coordinated and compliant.
Much of that review remains manual, and issues that escape it can become substantially harder to correct later. ArchiBoost supports architects, engineers, general contractors, and developers and owners in reviewing project information and identifying potential issues.
ArchiBoost is building systems that flag potential code compliance and constructability issues, discrepancies across disciplines such as architecture and mechanical, electrical, and plumbing, and conflicts between drawings and specifications, down to typos and inconsistent abbreviations. The goal is to amplify professional judgment and catch issues before they lead to requests for information, change orders, and costly rework.
In that role, Ku has contributed across product development and engineering. Through conversations with professionals across architecture, engineering, construction, and ownership (AECO), he gains insight into their workflows and challenges, helping the team translate those needs into product improvements and technical solutions.
“Being a founding engineer means helping drive the product, not just writing the code behind it,” Ku told TNGlobal in an interview. “You own much more of the loop between a customer’s problem and the product that eventually solves it.”
ArchiBoost’s approach centers on the interconnected information that AECO professionals work with every day: drawings, specifications, and project documents. Effective review requires preserving context across those sources, identifying potential issues, and providing supporting evidence that professionals can evaluate within their existing workflows. Ongoing engagement with AECO firms helps the team refine these capabilities around practical review needs.
As a key member of the engineering team, Ku helped translate those needs into technical systems that analyze project information and connect findings to their sources. The technical challenge extends beyond retrieving individual regulations to interpreting relationships across disciplines and documents.
Teaching AI to Reason About Complex Spatial Relationships
Doing so, however, is far more difficult than recognizing what appears on a single page, according to Ku.
“A construction document set is a network of interdependent drawings, schedules, details, and specifications,” he said. “It is filled with symbols, dimensions and notes, but each piece only tells you part of the story.”
Depending on its scale, a project can span hundreds or thousands of drawing sheets, accompanied by thousands of pages of specifications. The work is divided among architectural, structural, mechanical, electrical and plumbing teams, each contributing different information to the same building. An issue may therefore remain invisible on any single page and emerge only when information from several disciplines is connected.
“What makes it difficult is gathering the connected and relevant context needed to evaluate an issue,” Ku explained. “A value, symbol or tag may be defined in a different file or hundreds of pages away. A room detail may be one small part of a larger floor plan, while a row in a schedule somewhere else determines how it is meant to be built.”
The system must trace these references while determining which information matters and whether anything essential is missing. Retrieving too little can leave out a critical dependency, while processing everything at once can bury the relevant clue in noise.
Building-code compliance adds another layer. The governing requirement can change based on where a project is located, which code edition applies, and how the building will be used and constructed. Local amendments and qualifying exceptions can change the answer again. Finding a relevant passage is therefore only the beginning. The system must determine whether the rule applies to the project and which facts from the drawings support its conclusion.
Building Beyond the Model
Addressing this challenge requires more than a powerful AI model. ArchiBoost’s review workflows rely on an agentic engine that gives agents memory, instructions, and access to tools while coordinating how they work through a problem. Ku designed and helped build the agentic engine alongside the engineering team. He continues to contribute to its development.
“A powerful model is only one part of the solution,” Ku noted. “The engine allows it to break a problem down, follow new evidence and decide what it needs to investigate next, instead of simply producing an answer in one attempt.”
When examining a building-code question, for example, the engine can divide the problem into smaller tasks. One agent may interpret the relevant drawings while another researches the applicable regulations and other authoritative sources. The agents can use tools to search documents, examine images, or perform calculations, then carry what they learn into the next stage of the investigation.
Finding a relevant rule may create another question rather than end the search. If an agent discovers an exception, it must determine whether the project qualifies for it and gather the evidence needed to decide. If a drawing reveals a new condition, the engine can adjust its plan, retrieve additional context, or delegate a focused task to another agent.
Not every step calls for AI. Calculations, structured searches and clearly defined requirements may be handled more reliably through conventional rule-based software. The engine coordinates these different methods so that each is used where it is most appropriate, while preserving the project context, intermediate findings, and supporting sources gathered along the way.
Domain expertise is embedded in the core of ArchiBoost’s platform, shaping its system architecture, review methodology, and interpretation of complex project information. The engine performs its analysis autonomously, without human review or validation before findings are delivered.
Potential issues are presented as review tickets with locations and supporting evidence. Clients evaluate the findings and provide feedback that helps the team improve the system with each project.
“More autonomy is not always better,” Ku said. “If information is missing, a regulation is open to interpretation or a decision requires professional judgment, the system should not guess.”
Surfacing Potential Issues Before Construction
On a Massachusetts laboratory project already undergoing permit review, ArchiBoost flagged a non-potable cold-water line shown supplying an ice maker used for laboratory experiments.
Identifying the connection required interpreting the NCW abbreviation on a plumbing drawing and tracing the supply to the equipment it served.
The engine generated a review ticket linking the drawing location to a potential code issue, citing Massachusetts plumbing requirements governing potable-water supplies for specified uses, including the processing of medical or pharmaceutical products. This gave the client a concrete finding and supporting reference to evaluate against the equipment’s intended use.
Ku said the client told ArchiBoost that neither its project team nor its plumbing engineer had identified the connection. The client said the issue would most likely have surfaced during inspection and estimated that correcting it at that stage could have cost five figures, potentially reaching six figures.
For Ku, the case illustrated how the system could support professional review by connecting the evidence and presenting a potential issue for evaluation. Identifying the connection in the design documents gave the client an opportunity to determine whether a correction was needed before construction.
Building Companies Worth Trusting
The non-potable water case demonstrated what the system could do on one project. For Ku, however, the stronger measure of trust is what customers do afterward. Do they return with another project? Do they expand how they use the product? Do they examine its findings and act on them?
Ku said ArchiBoost began seeing those signals when one of its early customers repeatedly returned with real projects and asked how the platform could support more of her workflow. That engagement helped the team refine review workflows and identify additional ways the platform could address AECO firms’ day-to-day needs.
Trust, in this sense, does not require professionals to believe that AI will always be correct. It develops when a system fits into their workflow, produces useful findings, shows the evidence behind them, and remains clear about where professional judgment is still required.
Ku believed this distinction will become more important as powerful models become cheaper and more widely available. Access to AI alone will offer companies less of an advantage. The greater challenge will be turning that capability into products that understand a particular industry’s information, workflows and standards of accountability.
“As access to strong models becomes easier, the question shifts from ‘Who has access to AI?’ to ‘Who can turn that intelligence into something genuinely useful?'” Ku said.
That shift is also changing the role of the engineers building these products. As Ku sees it, their differentiating skill will increasingly be product taste: a form of judgment that is harder to quantify than model performance.
“The strongest AI engineers will increasingly look like product builders,” Ku explained. “They will still need deep technical ability, but they will also need judgment about users, workflows, design, trust and ultimately what is worth building in the first place.”
It is a lesson Ku expects to carry throughout his career, both as an engineer and in companies he hopes to build one day. His experience at ArchiBoost has reinforced his belief that a durable technology company cannot be built around technical capability alone. It must begin with an important problem, understand the people responsible for solving it, and earn a place in how they work.
For Ku, the companies that endure will not be those that ask professionals to trust AI simply because it appears intelligent. They will be those that build products worthy of that trust.