Research & Projects

As AI enters the factory floor, a new question is emerging: How should AI communicate with the people working alongside it?

Today, AI systems often rely on screens, alarms, and notifications to get workers’ attention. But more information does not necessarily mean better communication, it can add cognitive load, interrupt workflow, or even contribute to distrust and alarm fatigue. My PhD research explores a different possibility: What if AI could communicate through music? I developed a Musical Nudge framework that uses changes in tempo, pitch, and rhythm as a subtle communication channel between humans and AI. Rather than interrupting workers with explicit commands, musical nudges are designed to influence attention and behavior while allowing workers to remain in control. The research brings together Human–AI Teaming, behavioral science, music, and smart manufacturing, grounded in the Cyber-Physical-Social Systems (CPSS) perspective. This work was conducted at the University of Georgia with Dr. Beshoy Morkos and collaborators.

Objectives

I set out to understand whether music could become a meaningful communication channel for human–AI collaboration:

• Develop a framework for using musical cues as communication between humans and AI.

• Investigate how musical nudges influence trust, performance, workload, and behavior.

• Compare communication from human and AI sources to understand how the perceived source of a nudge affects workers' responses.

• Identify strategies for designing AI communication that supports appropriate trust, neither blind reliance nor unnecessary distrust.

To study these questions in a realistic setting, I designed and built a simulated manufacturing assembly line with four workstations and 21 assembly steps. The system combined behavioral and physiological measures, including EmotiBit biosensors, load cells and weight sensors, and synchronized video capture. I conducted a mixed-methods study with 20 engineering participants, examining task performance, workload, physiological responses, and participants' experiences of working with AI. Data were analyzed using Python, R, MAXQDA, and NASA-TLX, integrating quantitative performance and physiological measures with qualitative interview data.

Results & Impacts

In a pilot study with 20 participants, early results suggest that auditory nudges, designed to simulate AI communication, can shape how people work without adding detectable physical strain.

• Musical nudges sped people up. Participants tended to work faster when the AI guided them with musical cues.

• People’s response changed over time. As participants heard nudges again and again, their reactions shifted. The chime showed the clearest change, which suggests that people learn and adapt to musical cues.

• Nudges didn’t add stress. Heart rate and skin conductance stayed about the same with and without nudges, so the music guided people without making their bodies react as if under pressure.

• No single nudge design won. Different musical cues worked similarly well. This suggests that how and when a nudge is used may matter more than the exact sound. 

This research moves beyond asking whether AI can communicate with workers to a more fundamental question:

What should AI communication feel like when humans and machines work together?

The framework was published in Manufacturing Letters (2025) and presented at NAMRC 2025. Two follow-up papers, currently under review, examine (1) how auditory nudges affect performance, physiology, and workload, and (2) how workers adapt to, ignore, or escalate their response to them.

This work has contributed to two NSF proposals currently under review. Since 2022, my research on trust and human-AI communication has been recognized with the 2024 ASME/CIE Sci Rising Award and first place in two NSF/ASME Design Essay Competitions (2022, 2023).

Research Toolkit

Human-AI Teaming · Human Factors · Experimental Design · Mixed Methods · Biometric Analysis · Wearable Sensors (EmotiBit) · Python · R · MAXQDA · NASA-TLX

Pecan shelling still relies heavily on decades-old processes, and small inefficiencies add up to lost time, energy, and product quality. Our USDA-NIFA research program at the University of Georgia works with the pecan industry to modernize shelling through advanced manufacturing, from how pecans are prepared to how they are cracked and processed.

For new technology to work well, machines need to understand and adjust to the pecans they are processing. That starts with getting pecans into the right condition and understanding what "right" means. One of the most important conditions is moisture. Too dry, and kernels shatter into broken pieces. Too wet, and quality and efficiency suffer.

My research focuses on this step and asks, how can we reliably bring pecans to a target moisture level? I compare different conditioning methods, including conventional hot- and cold-water soaking and vacuum (negative pressure), and develop a "recipe" for each one. Each recipe shows how combinations of time, temperature, and pressure lead to a predictable moisture range. With these recipes, processors can choose the method and settings that fit their machines, their target moisture, and their priorities, whether that's speed, energy use, or kernel quality.

Each method comes with its own tradeoffs. Our goal is to map them clearly so the industry has a practical set of options instead of a one-size-fits-all approach.

Objectives

• Characterize how temperature, soaking time, and pressure affect moisture uptake in pecan kernels

• Build moisture "recipes" for conventional soaking and vacuum conditioning

• Evaluate tradeoffs between conditioning speed, energy use, and kernel quality

• Find practical, faster alternatives to the slow, destructive oven method for measuring moisture

• Translate lab findings into usable processing strategies for the pecan industry

Approach: I helped design and run a systematic study of more than 100 conditioning treatments, using Latin Hypercube Sampling and response-surface methods. We measured moisture with the standard oven method and compared it against several commercial moisture analyzers. I also trained and supervised 12 undergraduate researchers and worked closely with industry stakeholders throughout the project.

Results & Impact

• Each method has its own range. Time, temperature, and pressure combinations produced predictable moisture ranges, giving processors several routes to a target moisture level.

• Vacuum offers a fast option. Vacuum conditioning raised kernel moisture by about 2 percentage points within minutes, compared with hours for conventional soaking.

• There's an effective pressure window. The most effective vacuum range was about 17.5–22.5 kPa absolute pressure.

• Heat has tradeoffs. Higher temperatures sped up moisture transfer but caused discoloration and more variation between kernels.

• Faster moisture measurement. Among the analyzers tested, Steinlite was the most practical alternative to the oven method once a small bias was corrected.

This work gives the pecan industry a clearer map of its conditioning options, so processors can match their method to their equipment, target moisture, and priorities. It also expands the range of available methods to include much faster vacuum-based conditioning.

Findings on hot- and cold-water soaking were published in Food and Bioproducts Processing (2025), and the vacuum conditioning work is under review. I have presented results at ASME-IDETC (2023) and have led annual USDA workshops since 2023, sharing findings with 120+ industry partners and stakeholders.

Research Toolkit

Experimental Design · Process Optimization · Vacuum Processing · Latin Hypercube Sampling · Statistical Modeling · Measurement System Analysis · Python · MATLAB · Team Leadership · Industry Collaboration