Stay up to date on upcoming events, deadlines, news, and more by signing up for our newsletters!
At first, it was Pac-Man. Literally.
Pac-Man that gobbled up antibiotic complications.
At the Cornell hackathon where Madhu Nzerem (GSAS ‘26) and Dr. Khanh Pham first pitched their idea, they called it PACMAN-EHR — short for Preventing Antibiotic Complications in Electronic Health Records. Clips of the arcade game played in the background, but the problem they were addressing was anything but lighthearted: antibiotic errors are alarmingly common, even among seasoned physicians, and can lead to serious patient harm.
“In my daily routine of seeing patients with infectious diseases, I realized there was a need for an AI-powered assistant to act as a safety net for antibiotic prescribing,” said Pham, the company’s CEO, a physician at NewYork-Presbyterian, and assistant professor at Weill Cornell.
Now rebranded as StewardGuard, the tool has evolved into an EHR-integrated clinical decision support system designed to reduce errors in antibiotic use. It automates the manual review process by extracting key patient data from electronic health records and cross-referencing it with treatment guidelines. Combining software logic with AI assistance, StewardGuard recommends safe, empiric antibiotic options in real time.
“Physicians are strained for time and pulled in every direction,” said Nzerem, the company’s CTO and a PhD candidate in computational biology at NYU Langone. “StewardGuard eliminates the need to dig through patient records, especially in those first critical moments of care.”
To ensure safety and reliability, the software uses a two-tiered architecture. Most of its logic is hard-coded based on clinical guidelines, while large language models are employed selectively to validate inputs and handle ambiguity.
“We try to minimize the AI’s role,” Nzerem said. “Hallucination risk is real, especially in a clinical setting.”
In early tests with more than a dozen simulated inpatient cases, StewardGuard made antibiotic decisions 33 times faster than clinicians — with 100% adherence to established guidelines.
Speed alone, however, isn’t enough. Running large models locally requires GPU infrastructure that many hospitals lack.
To address this, the team is testing smaller models and exploring partnerships with OpenAI and other cloud providers to strike a balance between performance and data security.
Beyond technical development, StewardGuard has also taken steps to protect its intellectual property. Before joining NYU’s Summer Launchpad accelerator, the team drafted and submitted a provisional utility patent through Cornell’s technology transfer office to secure its proprietary code and clinical workflow. They hope to finalize the filing by the end of the summer, marking an important step toward commercialization.
With support from the Entrepreneurial Institute, the team is now in discussions with NYU Langone and Weill Cornell to prepare for potential in-hospital trials — a critical phase in their broader goal of deploying StewardGuard in smaller community hospitals, where physicians often lack access to infectious disease specialists.
“They might have the bigger ‘hair-on-fire’ problem,” Pham said. “We’re building this for the places that need it most.”
