You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$60,000 left to reach the funding goal.
Scientific research and technical assessments require substantial public and private investment. Producing reliable evidence, however, does not guarantee that decision-makers will interpret it correctly. A conclusion can survive translation while the conditions that made it valid do not.
In construction, life cycle assessments (LCAs) and environmental product declarations (EPDs) provide detailed information about building materials, manufacturing, and environmental performance. When these findings move from researchers to policymakers, procurement professionals, and government agencies, important context, including system boundaries, assumptions, and uncertainties, may be simplified or lost. AI can help process this information, but a fluent summary does not necessarily preserve the scientific conditions that give a conclusion its meaning.
Lost in Translation proposes EvidenceBridge AI, an experimental system designed to preserve and verify technical evidence as it is translated into information used for policy and procurement decisions.
The project's central innovation is combining three capabilities:
Evidence preservation and verification: Connect technical claims to their original sources, retain critical assumptions and uncertainties, and flag conclusions that exceed what the evidence supports.
Jurisdiction-aware interpretation: Compare technical findings with selected building regulations, procurement requirements, and standards in two U.S. jurisdictions, identifying where evidence is applicable, insufficient, or incompatible with specific requirements.
Decision-oriented communication: Generate understandable, source-traceable explanations that help nontechnical decision-makers interpret findings without losing essential qualifications. Where supported by the evidence, these explanations may also address implications for material selection, manufacturing, and domestic supply chains.
The research will test whether this approach performs better than conventional AI and human-written summaries. We will develop a human-reviewed reference set using published LCAs, EPDs, and policy documents, then evaluate all three approaches against the same predefined criteria.
Our primary measure will be evidence completeness, defined as the proportion of decision-relevant information preserved during translation. Additional measures will assess technical accuracy, source traceability, applicability, and unsupported claims.
The project will produce a documented AI prototype, a reproducible evaluation methodology, and openly available findings and supporting materials where licensing permits. We intend to share the results with relevant professional organizations and public agencies, including NYSERDA and the New York State Department of Environmental Conservation, without assuming their participation.
The goal is to establish whether AI can make complex scientific evidence more accessible without compromising the information necessary to interpret and use it responsibly.
The goal of Lost in Translation is to find out whether an AI system that keeps track of its sources can carry the meaning of building research into policy and procurement documents more accurately than a standard AI summary or a summary written by a person.
I am developing EvidenceBridge AI, an experimental tool. It links each technical claim to the place in the original study where it comes from, keeps the conditions that limit the claim, and checks whether the evidence applies to a specific requirement.
I will build a reference set from published life cycle assessments (LCAs), environmental product declarations (EPDs), and building-sector policy documents. Two independent reviewers will mark the information that matters for interpreting each source: the key numbers, the system boundaries, the assumptions, the uncertainties, the limitations, and the conditions under which the result applies. I will define these items before anyone scores a summary, so every summary is judged against the same list.
A claim counts as preserved when its number (where it has one), its system boundary, its uncertainty, and its conditions of applicability all remain accurate in the summary.
My main measure is evidence completeness: the share of those predefined items that survive into each summary.
To check that the reviewers agree with each other, I will report percent agreement and Cohen’s kappa. My working target is 0.70 for categorical ratings. If agreement is lower, I will tighten the rubric, have both reviewers rescore a subset, and record any disagreements that remain before I finalize the reference set.
I am building a limited research prototype with three functions:
Source tracing: linking each claim to a specific location in its source and writing an accessible explanation that keeps the important qualifications.
Evidence checking: flagging conclusions the data does not support, missing assumptions, comparisons that do not match, and uncertainties that could change the answer.
Jurisdiction checking: comparing the evidence with selected regulations, standards, and procurement requirements in two U.S. jurisdictions.
For the jurisdiction check, two reviewers will decide whether the evidence satisfies, partly satisfies, or does not satisfy each requirement. If the documents are not enough to decide, the item will be marked not determinable. I will set these judgments from the governing documents before I look at any AI output.
I will then compare EvidenceBridge AI with human-written and standard AI summaries, using the same reference set and the same criteria. Evaluators will score each output without knowing which method produced it. I will keep records of the model names, versions, prompts, and evaluation steps so that someone else can repeat the work.
Evidence completeness is my primary result. I will also report numerical accuracy, source tracing, applicability, and unsupported claims, along with the uncertainty around each difference.
I will write material selection and procurement scenarios based on the published requirements of the two jurisdictions. Each scenario will state the decision, the available technical evidence, and the rules that apply. For each translated output, I will check whether it contains what a reader needs to know what the evidence supports, what remains uncertain, and what documents are still missing.
I will also look at whether the EPDs and procurement rules separate domestic products from imported ones, by manufacturing location, geographic coverage, or sourcing requirements. Where the documents do not answer the question, I will say so rather than guess.
I plan to share the findings and the prototype with professional organizations and public agencies, including NYSERDA and the New York State Department of Environmental Conservation. I am not assuming they will take part.
A later, optional study could look at how the summaries affect how nontechnical readers understand a decision. I would only run it if the responsible institutional review board approves it or formally determines it exempt.
At the end of the project I will have a working research prototype, a method that others can repeat, and public findings and materials, wherever the data licenses allow.
I am not assuming EvidenceBridge AI will perform better. The question is whether it preserves more of the evidence, under what conditions it fails, and where a person still needs to check the work.
I want to find out whether EvidenceBridge AI can explain complex building research more clearly without losing the information that makes the findings reliable.
The main thing I will measure is how much important information survives when technical research is translated into language that policymakers and procurement professionals can understand. This includes numerical results, assumptions, system boundaries, uncertainties, and limitations.
I will evaluate five things:
Completeness: How much of the important information from the original research is preserved?
Accuracy: Are the numbers and technical findings represented correctly?
Traceability: Can readers check where each important claim came from?
Applicability: Does the explanation correctly identify whether the evidence satisfies, partially satisfies, or does not satisfy a particular policy or procurement requirement, or whether there is not enough information to determine this?
Unsupported claims: Does the explanation introduce conclusions that the original research does not support?
I plan to start with approximately 30 published technical documents. Each document will be evaluated using three approaches: a human-written summary, a conventional AI summary, and an EvidenceBridge AI summary. Using the same documents will allow me to make direct comparisons. I will determine the final sample size through a statistical power or precision analysis before testing.
Two independent reviewers will help establish which information must be preserved. I will use a predefined scoring system, measure how consistently the reviewers agree, and aim for a Cohen's kappa of at least 0.70 where appropriate. If agreement is lower, I will clarify the criteria and repeat part of the review. The summaries will then be evaluated without revealing which approach produced them, and I will report the differences along with their uncertainty.
I will consider the project completed when I have developed a working prototype, tested it against the other approaches, and published the methodology and findings.
However, having a working tool does not automatically mean it is effective. EvidenceBridge AI will be considered effective only if it preserves more important information without introducing serious errors, such as reversing a numerical comparison or incorrectly stating that a building material meets a regulatory requirement. I will define the full evaluation criteria before testing.
I also recognize that EvidenceBridge AI may not perform better than existing approaches. If that happens, I will report it. Understanding where AI makes numerical mistakes, overlooks technical limitations, misinterprets policy requirements, or loses important context is valuable in itself.
Ultimately, I want to understand whether AI can help scientific evidence reach decision-makers without changing what the research actually says, and where human expertise remains necessary.
I am requesting $60,000 over approximately 12 months to develop and evaluate EvidenceBridge AI, with a minimum of $15,000 to begin the work. I will lead the project, drawing on my background in architecture, life cycle assessment, and sustainable construction.
Full budget: $60,000
Research time and project management: $24,000
This covers my work designing the methodology, building the reference set with the reviewers, developing and evaluating the prototype, analyzing the results, and writing the public report.
EvidenceBridge AI development: $12,000
This covers outside technical help, such as a software developer, to build the prototype. It includes document processing, linking claims to their original sources, flagging missing or unsupported information, and comparing evidence with the requirements of two U.S. jurisdictions.
Independent technical evaluation: $8,000
Two independent reviewers will build the reference set for about 30 source documents, score the outputs, and resolve disagreements. The estimate is 160 hours at $50 per hour, which I will verify before submission.
Human-written comparison summaries: $4,000
Independently prepared summaries of the same about 30 documents, providing the baseline against which EvidenceBridge AI and conventional AI are compared.
AI models and computing: $6,000
Access to AI models, computing, document processing, and repeated runs needed to measure how consistent the outputs are.
Statistical analysis and consultation: $3,000
Independent statistical advice on reviewer agreement, uncertainty estimates, and analysis software.
Public dissemination: $3,000
Publishing the methodology, scoring rubric, prompts, results, and any materials whose licenses permit sharing. This also covers sharing findings with researchers, building professionals, and public agencies, including NYSERDA and the New York State Department of Environmental Conservation. Their participation is not assumed.
If only $10,000 is raised
With $10,000, I will complete the scientific foundation of the project rather than the full prototype. That phase includes:
Researcher time: $4,000
Independent reviewers, building the reference set for a smaller sample: $2,000
Human-written summaries: $2,000
AI testing and a preliminary prototype: $1,000
Statistical consultation: $1,000
This phase will produce the reference set and scoring rubric, a preliminary comparison of human and conventional AI summaries, and an early test of the evidence-aware approach. Findings from this phase will be published and will determine whether the full prototype and the larger evaluation are worth pursuing. The complete EvidenceBridge AI prototype and the full evaluation will depend on additional funding.
I am Ivana Krsteska, a licensed architect, Fulbright alumna, and PhD candidate in Sustainable Construction Management at the SUNY College of Environmental Science and Forestry. I hold two master’s degrees in architecture and have worked in building research, environmental assessment, AI evaluation, and European Commission project evaluation.
I will lead EvidenceBridge AI. I have not yet recruited a formal team. My background is in architecture, life cycle assessment, and research methodology. Project funds would pay for specialized software development and independent technical reviewers.
Previous work
Northern Border Regional Commission (NBRC): I worked with three professors on a study of six mass timber buildings using comparative life cycle assessment. I contributed to the analysis of material performance and life cycle impacts, and to communication with industry and state stakeholders. The work showed me how technical results are interpreted once they reach people making decisions about construction, manufacturing, and regional development. [Confirm my specific role and the study’s outcome.]
Oak Ridge National Laboratory (ORNL): In 2025 I worked as a research intern on the U.S. Department of Energy’s Weatherization Assistance Program evaluation. I helped develop methods for measuring benefits beyond energy savings, including indoor environmental quality, health, and resilience. I worked with anonymized datasets, reviewed data quality, and supported technical reporting.
European Commission: From 2021 to 2025 I served as an external project evaluator for Erasmus+ and European Solidarity Corps programs. I assessed projects on quality, implementation, compliance, and results, and wrote evaluations with recommendations. [Number of projects evaluated.]
Publications and manuscripts
Determining the Sustainability of Mass Timber Structures Through Third Party Sustainability Assessment Systems. Associated Schools of Construction Proceedings, 2024, with A. Mirando and L. Onsarigo.
Human in the Loop, Not Hands Off: Do AI Assisted TA Comments Improve Construction Writing Quality and Fairness in an Introductory BIM Course? Associated Schools of Construction Proceedings, 2026, with Baowen Zhang and Paul Crovella. [Confirm the author list.]
Green Skin Architecture, a manuscript on the design and feasibility of a prefabricated bio based façade using moss, hempcrete, and a ventilated rainscreen. Under revision for [journal].
AI evaluation
I have worked as an AI domain expert evaluator, writing technical prompts and checking AI generated answers for numerical accuracy, unsupported assumptions, and logical inconsistencies. [Confirm that you can describe this work publicly under your contract.]
Credentials and teaching
I received a Fulbright scholarship for graduate study at Kent State University. I placed third nationally in the U.S. Department of Energy’s 2024 JUMP into STEM building science competition. I hold a LEED Green Associate (v5) credential and the Sustainability Excellence Associate certification. I have taught building information modeling, sustainable construction, building performance, and engineering ethics. [Confirm whether these were courses you taught or teaching support roles.]
Presentations
In October 2026 I led a roundtable at the Fulbright Association Annual Conference titled “From Data to Decision: How Researchers Can Bridge the Gap Between Technical Evidence and Policy Action.”
Why this project
At ORNL, I saw how outcomes that matter to people are often missed by standard metrics. Through construction research, I saw that researchers, manufacturers, and policymakers ask different questions of the same technical information. Through AI evaluation, I saw how easily a fluent answer can drift from the evidence behind it. EvidenceBridge AI investigates what survives when technical research reaches the people who make decisions, and what gets lost along the way.
I will recruit two independent reviewers with relevant credentials before the reference set begins, and I will record their qualifications. I will use a predefined rubric and measure their agreement. I will revise unclear criteria before the reference set is final and before any AI output is scored. If disagreement remains, I will report it rather than present the results as more certain than they are. If qualified reviewers cannot be recruited in time, I will not score the outputs myself without disclosing that limitation.
Commercial AI models are updated regularly, and the same prompt can produce different answers. I will record model names, versions, dates, settings, and prompts for every run. I will run all three approaches within a testing period fixed before testing begins. If a model changes significantly during that period, I will document the change and repeat the affected tests where possible.
LCA reports, EPDs, and government requirements use different formats, terminology, and reporting rules, so connecting them accurately may take more time and expertise than planned. I will start with selected document types and two U.S. jurisdictions. If development falls behind, I will complete the reference set and the baseline comparison first, since neither depends on the prototype being finished.
Certain EPDs, reports, or project documents may have licensing limits, and others may lack the information needed for a reliable comparison. I will verify permissions before testing and prioritize publicly accessible documents. Where a document cannot be redistributed, I will publish its identifier, the reason it is restricted, and the limits this creates for reproducibility.
The prototype could preserve no more information than conventional AI or human written summaries. It could also improve completeness while introducing errors in numerical interpretation or applicability judgments. I will report these results with uncertainty estimates and identify which kinds of information were most often lost or misread. A prototype that does not outperform the baselines would still show where AI assisted interpretation is unreliable, which is useful to anyone considering these tools for research or decision support.
The most serious failure would be building a tool without a reliable way to judge its outputs. I will avoid that by fixing the evaluation method before completing the prototype, and I will report the results whether or not they match my expectations.
I have not raised any grant or external research funding in the past 12 months. My previous doctoral funding arrangement did not continue, and I am currently developing this project independently while completing my PhD.