How Chandana Dayapule Applied AI to Contractor Prequalification and Risk Assessment

For artificial intelligence, extracting information from a document is becoming increasingly routine. Understanding what that information means in the context of a consequential business decision is considerably harder. That distinction became central to Chandana Dayapule's work at ConstructionBevy, where she applied machine learning, document intelligence, and structured decision systems to contractor prequalification and risk assessment.

Construction presents a particularly difficult information problem. Before awarding work, organizations may need to evaluate contractors using financial statements, compliance records, operational histories, qualification documents, and information collected from multiple internal and external sources. The difficulty is not simply that the information is distributed. It can be incomplete, inconsistent, outdated, or recorded in different formats, while the significance of a particular data point may change depending on the contractor, the project, and the organization making the decision.

The technical challenge, therefore, extends well beyond digitizing paperwork. It is about transforming fragmented evidence into a structured representation that can support a defensible decision. For Dayapule, the work represented another version of a research problem that had appeared throughout her career: how should an intelligent system make useful decisions when the necessary evidence is distributed across multiple sources, and uncertainty cannot simply be eliminated?

The Intelligence Problem Starts Before the Model

Many discussions of artificial intelligence begin with the algorithm. Contractor prequalification begins much earlier, with the information that reaches it.

Financial statements contain one category of evidence. Compliance documents contain another. Operational histories, historical records, and external sources contribute additional signals. Before predictive or risk models can be useful, those inputs need to be identified, interpreted, normalized, validated, and connected to the correct contractor and reporting period.

Dayapule's work approached this as an intelligence pipeline rather than as an isolated modeling problem. Document-processing systems could identify relevant information from unstructured records. Extraction mechanisms could transform that information into structured fields. Normalization could create consistency across documents and sources, while validation mechanisms could compare values, identify inconsistencies, and determine when information required additional review.

The objective was to create structured contractor information that could support subsequent analysis instead of forcing each evaluation to begin again from a collection of disconnected documents. The distinction is important because even a sophisticated predictive model cannot compensate for unreliable information entering the system. A contractor may be associated with an incorrect document, a financial figure may refer to the wrong period, two sources may contain contradictory values, or an important field may be missing altogether.

In each case, the eventual quality of the assessment depends on decisions made before the predictive model is invoked. One of the broader lessons of production artificial intelligence is that reliability begins with the information architecture surrounding the model, not simply with the final prediction.

Turning Documents Into Decision-Ready Information

Extracting a number from a document is not the same as understanding it. A financial value becomes useful only when the system knows what the number represents, which organization and reporting period it belongs to, how it relates to other available information, and whether the underlying source should be considered reliable.

This is where provenance becomes important. A useful decision-support system should preserve the relationship between a structured data point and the evidence from which it originated. Without that traceability, automation can make a process faster while simultaneously making its conclusions more difficult to inspect.

Dayapule's work emphasized moving from isolated records toward structured contractor information that could be validated, updated, and reused. That represented a shift from document processing toward decision-ready data. Rather than treating each financial statement, compliance record, or operational document as an independent artifact, the system could use those sources to build a more coherent representation of a contractor.

The architecture reflects a broader challenge now appearing across enterprise AI. Large language models have dramatically improved the ability of machines to interpret unstructured information, but enterprise decisions usually require more than linguistic understanding. They also require consistency, validation, provenance, recency, and confidence in the source. A system may be capable of reading a document fluently while still being unable to determine whether the information it extracted should materially influence a decision.

That gap between understanding information and establishing whether it can be trusted is one of the central problems production AI systems increasingly have to solve.

Why Contractor Risk Cannot Be Reduced to One Universal Score

Once contractor information has been structured, another challenge emerges: what exactly constitutes risk?

There is rarely one universal answer. One organization may place greater emphasis on financial resilience, while another may care more about operational capacity, compliance, project history, or other risk dimensions. Even within the same organization, the importance of individual factors may change according to project characteristics, contract size, scope, or internal policy.

A fixed score can therefore conceal an important assumption—that every organization evaluates every contractor according to the same definition of risk. Dayapule's work explored a more configurable approach in which multiple categories of evidence could be evaluated while assessment criteria reflected the context in which a decision was being made.

Conceptually, this separates the problem into two layers. The first is concerned with determining what the available evidence says about the contractor. The second concerns how that evidence should be interpreted for a particular decision.

That separation is significant because two organizations can review the same contractor information and reasonably reach different conclusions. They may operate under different policies, different tolerances for risk, and different project constraints. An intelligent risk system therefore has to represent not only the subject being evaluated, but also the decision context in which that evaluation is taking place.

The idea extends well beyond construction. Credit assessment, insurance, procurement, cybersecurity, compliance, and many other enterprise domains face a similar problem: the underlying data may be comparable, but its significance changes depending on who is making the decision and under what conditions.

AI as Decision Support, Not an Oracle

The purpose of a system like this is not to eliminate professional judgment. It is to make professional judgment better informed.

Construction risk decisions frequently involve situations that are difficult to compress into one automated answer. Information can be missing. Evidence can conflict. A contractor may not resemble examples represented in historical data. External conditions can change, and individual projects can create requirements that a generalized model does not fully capture.

A well-designed AI system should make those conditions more visible rather than hiding them behind a numerical score. That is why explainability and traceability become especially important in risk-oriented applications. A user should be able to understand which dimensions of the available information contributed to an assessment, what evidence supported those conclusions, and where uncertainty remains.

This is also where artificial intelligence begins to differ from simple workflow automation. Moving documents between systems can make a process faster. Extracting relevant evidence, reconciling inconsistencies, evaluating information in context, and presenting that reasoning to a human decision-maker begins to change the quality of the decision itself.

In this sense, AI does not replace judgment. It reorganizes information so that judgment can operate with a clearer and more complete view of the available evidence.

The Model Is Only One Layer of the System

Dayapule's work at ConstructionBevy also illustrates a broader principle that has become increasingly important in contemporary artificial intelligence: a production model rarely operates alone.

Before an output appears, information may have been collected, retrieved, extracted, transformed, normalized, validated, and combined with information from other systems. After the model produces a result, that output may be compared with additional evidence, evaluated against organizational rules, presented to a user, or incorporated into a larger workflow.

"The model can be very capable and still not be the whole system," Dayapule said. "You have to think about what happens before and after the model produces an answer."

Contractor risk makes that systems perspective especially concrete. A better predictive algorithm cannot correct a document that has been associated with the wrong contractor. A more capable language model cannot reason over information that was never retrieved. A numerical assessment has limited value when users cannot determine what produced it.

Reliability therefore emerges from the interaction among multiple layers: the information source, the extraction mechanism, the structured representation, the validation process, the model, the decision logic, and the human interface. The system succeeds only when those layers work together.

From Construction Intelligence to Modern Enterprise AI

The technologies surrounding artificial intelligence have changed rapidly, but many of the underlying systems problems remain remarkably consistent.

Retrieval-Augmented Generation, for example, allows language models to access external information that was not encoded during training. Yet a useful RAG system still has to answer many of the same questions that appear in document-based risk assessment. Which source should be trusted? Is the information current? Does one record contradict another? Is the retrieved evidence actually relevant? Can the system identify where its conclusion came from? What should happen when the available evidence is insufficient?

These are not secondary engineering details. Once an AI system depends on external evidence, information selection, provenance, and validation become part of the reasoning process itself.

Agentic AI makes those questions even more consequential. An agent may retrieve information, interpret it, choose a tool, execute an action, observe the result, and then use that result to determine another action. If one stage is unreliable, the error can propagate through the remainder of the workflow.

The lessons from contractor prequalification are therefore increasingly relevant to modern AI architectures. Grounding, provenance, validation, context, and explainability cannot simply be added after a model has been deployed. They have to be designed into the system from the beginning.

The Broader Research Problem: Reasoning Over Heterogeneous Evidence

The construction use case also points toward a deeper machine-learning problem. Real decisions rarely depend on one clean dataset. They depend on heterogeneous evidence.

Some information is numerical. Some is categorical. Some is written in natural language. Some resides in structured databases. Some is historical, while other information may be recent. Certain sources may be highly reliable, while others may be incomplete or uncertain.

An intelligent system therefore has to do more than produce a prediction from a fixed set of features. It must reason over evidence with different levels of reliability, freshness, completeness, and relevance.

That makes the problem as much about information architecture as about machine learning. Before asking whether a model can predict risk, researchers and engineers increasingly have to ask what evidence the system should observe, how that evidence should be represented, how contradictory information should be reconciled, which sources should receive greater weight, how uncertainty should be expressed, and how a human reviewer should understand why the system reached a particular conclusion.

Those questions have become increasingly important as artificial intelligence moves deeper into consequential enterprise decision-making.

Building AI for Decisions That Matter

Construction may appear far removed from the race to build increasingly capable foundation models, but the underlying problem Dayapule encountered at ConstructionBevy is becoming more common across industries.

Organizations possess enormous quantities of information, yet much of it remains fragmented across documents, databases, workflows, and external systems. Different decision-makers may apply different policies to the same evidence, and consequential decisions require something more rigorous than a plausible generated answer.

They require systems capable of organizing information, preserving its provenance, understanding context, exposing uncertainty, and supporting human judgment.

Solving that problem requires more than improving model accuracy. It requires designing the architecture surrounding the model: how information enters the system, how it is transformed into usable evidence, how conflicting information is handled, how context affects interpretation, and how conclusions remain traceable to the sources that produced them.

That is what makes contractor prequalification and risk assessment an instructive artificial-intelligence problem. The difficult part is not generating another risk score. It is building a system in which documents, structured data, machine-learning models, organizational context, and professional judgment work together well enough to make the assessment meaningful.

As AI moves further into decisions involving financial, operational, and organizational risk, that systems-level challenge may ultimately matter more than the sophistication of any individual algorithm.