MaryAdelua.

My Yorùbá work asks whether generated speech preserves the distinctions that a fluent listener needs to understand an utterance. This is a natural language processing (NLP) problem grounded in the language itself: lexical tone, vowel distinctions and diacritics carry information that a fluent-sounding voice can still lose.

My contribution

I develop a workflow that brings together natural-speech references, structured listening judgments, acoustic measurements and reviewable annotations. I use the results to identify the next measurement or experiment, including cases where an intervention does not help.

Listening and acoustic evidence

I distinguish lexical tone from broader prosody, including rhythm and intonation. I also examine segmental pronunciation—the consonants and vowels—alongside transcript match, intelligibility, meaning recovery and naturalness. These questions stay separate so that a natural-sounding recording does not automatically receive credit for saying the intended words.

The acoustic workflow extracts fundamental frequency (F0, a measure related to pitch), voicing probability and RMS energy with pYIN, and compares pitch estimates with Praat/Parselmouth. Speaker-relative pitch representations help describe contours; missing values and disagreements between estimators remain visible. A pitch contour is evidence for inspection, not an automatic verdict on correct tone.

Language-aware preparation and independent review

The text pipeline preserves Yorùbá tone marks and underdots through Unicode normalization. Automatic alignment suggestions and human-verified timing remain separate. The independent boundary-annotation tool is designed for reviewers to mark syllable and tone-bearing-unit intervals, vowel or syllabic-nasal nuclei, and uncertainty without seeing one another’s annotations or prior scores. Its JSON exports preserve each reviewer’s evidence. Building the tool does not itself establish reviewer agreement or complete the annotation study.

What the pilot can establish

The listening results are small, single-rater, unblinded diagnostic findings. A matched prompt-guidance intervention produced no aggregate gain in tone correctness or naturalness. Whole-utterance pitch similarity is not lexical-tone accuracy, and these results do not establish provider rankings or population-level performance. Further model-adaptation training remains pending.

My Smart pH project explores how a vaginal pH product concept could communicate measurements, uncertainty and next steps clearly. The current software is a research prototype using fictional readings and a mock detachable reader. It lets me work through the interaction before making claims about a physical product.

My contribution and current methods

I develop the concept and simulated app workflow around understandable information and explicit device states. The prototype distinguishes disconnected, calibration-required, ready and error states. Simulated invalid samples are rejected rather than stored or plotted, and users can review or delete locally saved fictional readings.

Four separate fictional scenarios explore an unusable sample, a change worth checking, an ordinary example reading, and symptoms despite an ordinary reading. Each separates what happened, what the example means, and a possible next step. The selected story determines the message; no clinical alert is inferred from a generated number. This is a way to examine the design of information and uncertainty, not evidence from a completed user study.

Separating software checks from measurement evidence

Automated checks cover the simulated reader’s states, invalid-sample handling, data storage and deletion, and the separation of fictional scenarios from reading history. A proposed laboratory feasibility plan would use buffer solutions and a reference instrument to investigate calibration, repeatability, drift and the effect of detaching the reader. Those are planned experiments, not achieved performance results.

Current limits

No working sensor performance, clinical screening or diagnostic accuracy has been established. The prototype contains no real patient measurements or human-testing results; it does not classify a person as healthy or unhealthy. Clinical interpretation and real-data workflows would require separate expert review and validation.

My cholangiocarcinoma research asks a focused evaluation question: does dividing hypothesis generation among specialized agents produce better-supported early-detection research ideas than a single language model with the same retrieval tools? The output is a set of hypotheses for investigation, not a cancer-detection system.

My contribution and comparison design

My work evaluates the single-model and multi-agent outputs using retrieved sources and structured research traces. The baseline proposes hypotheses directly. The multi-agent system separates literature, molecular and signal generation, then uses adversarial review and ranking. Shared model settings and retrieval access help make the comparison about the system architecture rather than unequal tools.

What gets evaluated

I examine whether cited sources exist, whether they support the specific claim, whether the hypothesis withstands an independent adversarial search, how novel it is, and how many valid supporting sources ground it. These are separate dimensions: a real citation can still fail to support a claim. The evaluation harness scores hypotheses without exposing their system of origin and retains structured outputs for inspection.

The recorded larger pilot evaluated 52 baseline hypotheses and 50 multi-agent hypotheses selected from 79 candidates. The results are mixed: the multi-agent outputs had greater supporting-source depth, while citation-support intervals overlapped and neither system’s hypotheses fully survived the adversarial evaluation. I use that pattern to identify limitations and further tests, not to claim that more agents are generally better.

Current limits

This is an early computational pilot with automated evaluation and live retrieval, not clinical validation. It contains no patient-data analysis and does not establish an effective biomarker, diagnostic model or improvement in patient outcomes. Small samples and incomplete early-run provenance also limit what can be concluded.

My human–computer interaction (HCI) focus is on the relationship between a tool and the person using it: their goals, interpretation, trust and decisions. I connect this focus to my background in interviews, surveys, qualitative and quantitative research, and thematic and content analysis. Each project’s design and stage determine which methods are actually used.

Workplace advice and support

In Asking the Machine, with Dr. Ioana Cionea, I study how employees use generative AI chatbots for workplace advice, support and decision-making. My work connects communication questions to human–AI interaction: what people seek from a chatbot and how it enters their working lives. The project is at the IRB stage; I am not presenting completed participant findings here.

Frontline technology

The planned American Frontline Technology Study 2027 asks about exposure to and adoption of emerging technologies among people in physical and non-office roles. It extends my interest beyond office-centered accounts of AI use. This remains a study direction, with no completed fieldwork or findings claimed.

From research questions to design

In my work at Gazillion CRI, I translate questions about AI-mediated communication and communication friction into evaluation plans and product requirements. Across these projects, I want to understand both task performance and the user’s experience of the interaction. A functioning prototype and a completed human evaluation are different stages of that work.

My planned Cross-Cultural AI vs. Human Survey Respondents study asks whether AI-generated responses can approximate human survey responses across cultural contexts. I am interested in where the resemblance holds, where it breaks, and whose experience is poorly represented.

The methodological direction is to compare simulated and human responses to the same questions, examine response patterns across contexts, and document disagreements as well as aggregate similarity. Such a comparison would need an appropriate human reference, clearly defined populations and questions, and evaluation beyond the examples used to develop the simulation.

This is a proposed direction, not a completed experiment. I do not yet claim a validated user simulator, evidence that synthetic respondents can replace participants, or reinforcement-learning experiments.

I want to connect computational methods with questions about how people interpret language, evaluate advice and make decisions. The connection to my current work is concrete: speech evaluation asks whether a listener recovers intended meaning, while workplace AI research asks how people use a machine’s response in a human setting.

One conceptual strand explores how organizational communication theory might inform candidate objectives and evaluation measures for language models. A future cognitive-science direction would turn a clearly stated explanation of human behavior into testable predictions, then compare those predictions with human observations and simpler alternatives.

I treat this as an emerging research agenda. I have not established a cognitive model, trained a model on these proposed objectives, or shown that an LLM reproduces human cognition. Similar answers alone would not establish the same underlying process.