In this special episode of On Record's Inside the VALT series, hosts Mike Anzalone and Carmen Haak speak with Eli Lipasti, Director of IT at Intelligent Video Solutions (IVS). Lipasti studied computer engineering at the University of Wisconsin-Madison and joined IVS right out of college as a support engineer, providing technical support for the VALT software and helping build new processes for the support team. His role since then has expanded to include product management and technical design work that ties customer feedback directly to new VALT development. He also takes part in ongoing conversations with prospective and current customers about their technical needs. Lipasti explains that IVS's approach to AI in VALT started with conversations, not a single attention-getting feature. Rather than building something flashy for its own sake, the team spoke with simulation instructors and faculty about where they needed help, and the common request was for tools that let a smaller team support a growing number of students. That feedback shaped the first feature: on-server transcription. IVS is adding a GPU to its appliance servers so that video and audio can be transcribed entirely on-site and offline, without sending data to an outside service. Lipasti ties this directly to privacy requirements common among IVS clients, including HIPAA and government-level compliance standards that many commercial transcription services do not meet. The next layer of features builds on that local transcription. Diarization will identify which participant is speaking at any given moment, so instructors can track how an individual student's communication changes across a program rather than only within a single recording. Automated summaries will highlight key moments, quotes, and action items within long recordings, reducing the time staff spend reviewing footage that runs far longer than anyone has time to watch in full. Lipasti notes that these tools are meant to give staff their time back, not to replace an instructor's judgment. Because VALT already lets teams tag and search recordings by participant, custom field, or sharing status, adding searchable transcript content to that same system means instructors can find a specific phrase or interaction across dozens of past recordings in seconds rather than hours. Model selection is its own area of ongoing work. VALT's current transcription uses OpenAI's Whisper, a widely used general purpose speech recognition model, but Lipasti notes that general models perform less well on medical terminology, drug names, and clinical procedure names. IVS is tracking a newly introduced benchmark called medical word error rate, which measures transcription accuracy specifically on medicine-related language, and plans to add a medicine-specific speech model for its next version alongside diarization and summaries. Looking further out, IVS is adding optical character recognition that reads vitals monitors during a recording, tracking blood oxygen, heart rate, and respiration rate as data points plotted alongside the video timeline, along with a customizable field for tracking any other number that appears on screen. Because the feature reads the vitals display rather than integrating with a specific manikin, it keeps VALT usable across manikin brands, an approach Haak notes is important to clients who use equipment from multiple vendors. Sentiment analysis is also part of the plan: assessing tone, pacing, and vocal cues in a recording rather than only the words themselves, both live and after the fact. Lipasti flags that state-level regulation of sentiment analysis is still developing, and stresses that the goal is to support a clinician's or instructor's own judgment rather than produce a diagnosis. Asked for advice for listeners considering AI tools of their own, Lipasti recommends starting with a clear-eyed look at data privacy needs, since that decision narrows which tools are realistically available. He also cautions against subscribing to too many AI services or add-ons at once, since managing a large number of vendor agreements and data policies can become its own burden, and some newer AI vendors may not stay in business long enough to be a safe long-term choice.
