Frontier open-weight AI models now rival commercial systems on many research and instructional tasks, but at 400 billion to 3 trillion parameters, they are larger than any single Utah campus can practically host. Today, a student's access to frontier AI depends on which institution they attend and which vendor contracts it can afford. The USHE [Utah System of Higher Education] Open Model Commons closes that gap. We will deploy a curated catalog of frontier open-weight models on Redtail and operate it as a single governed, OpenAI-compatible inference service available to every USHE institution through the AI gateways campuses already run. UVU's gateway is in production for all employees today, SLCC's is ready to launch, and several sister institutions are close behind. Integration is configuration, not new campus infrastructure. A two-tier quota model enforces hard, auditable fairness across institutions while each campus governs its own users, courses, and labs. Prompts and outputs stay inside Utah's public education network, are never used to train vendor models, and remain under institutional governance. Our six-month milestone: the service live at all eight degree-granting institutions, with courses and research groups onboarded and utilization published continuously. The project produces durable artifacts: a model catalog, an integration and governance playbook, and a transition plan into Redtail's quarterly allocation process. The Open Model Commons makes Redtail's value visible to every campus, and to tens of thousands of students, in its first operating period.

Epilepsy affects approximately 50 million people worldwide, and about one third of patients have drug-resistant epilepsy. Surgical resection can be highly effective when the epileptogenic zone is accurately localized, but presurgical evaluation remains uneven across epilepsy centers because expertise, imaging protocols, computational resources, and software tools vary substantially. This project proposes an AI challenge to accelerate EPIDEA: Epilepsy Platform for Informed Decision Making with Enhanced Artificial Intelligence, an open-source platform for epilepsy presurgical decision-making. EPIDEA integrates scalp EEG, intracranial EEG, MRI, and diffusion tensor imaging to support two linked tasks: interpretable focal epileptogenic-zone localization and multimodal epileptic brain network modeling. The Redtail AI Factory would uniquely enable this project by providing H200 GPU capacity, high-speed interconnects, and large scratch storage needed for scalable training, benchmarking, and optimization of deep learning models for EEG source imaging, biomarker detection, multimodal fusion, and brain network reconstruction. The near-term milestone is to complete a six-month pilot benchmark of the EPIDEA computational pipeline using public, synthetic, and de-identified research datasets. Expected outputs include a Redtail-compatible containerized workflow, benchmark results comparing AI-based and classical methods, initial network-modeling results, student training materials, and a manuscript/grant-ready technical report. The project will advance responsible biomedical AI, reduce disparities in access to presurgical decision-support tools for Utah patients, particularly in rural and underserved areas of the state, and train Utah students in high-performance AI for healthcare.

This project will advance environmental health research through development of innovative multi-exposure characterization tools leading to the creation of daily exposomic patterns for Utah within and across domains (e.g. climate, air pollution, social disadvantage) and made available to other researchers. Advances in generative artificial intelligence (AI) models, which at their core learn patterns in data to propose what comes next, have created a promising foundation for advances in constructing exposomic measures for environmental research. Our team has developed and implemented tools for generating and describing localized (state, multi-county) spatiotemporal exposomic profiles for specific domains (meteorological variables, industrial air pollutants). However, our work has been limited by processing and storage constraints. We will utilize the unique capabilities of Redtail and its advanced computing resources to accelerate this work, creating an agnostic description of the environment at scale that can be used to search for associations across disciplines and outcomes by: (1) developing AI tools to characterize and describe multi-domain exposome measures for Utah; (2) linking these measures to synthetic populations with hourly activity schedules to generate population-based exposomic profiles; and (3) characterizing uncertainty in exposomic measure assignment due to the "mobility" of the synthetic population. This will include tools to aid in interpretability and visualization. This project aligns with several national initiatives, including the National Institutes of Health (NIH) Health & Extreme Weather (HEW) Data Accelerator, and the National Artificial Intelligence Research Resource (NAIRR) Pilot, and will provide foundational data and methods for these external funding sources.

Per- and polyfluoroalkyl substances (PFAS) pose a severe, escalating threat to Utah's groundwater/drinking water (such as Salt Lake City, Park City, etc.) and agricultural ecosystems. Due to their exceptionally strong C–F bonds, traditional water treatments fail to destroy these "forever chemicals," necessitating the development of novel catalytic materials, such as High-Entropy Alloys (HEAs) and transition bimetallics. However, the combinatorial explosion of potential catalyst compositions makes empirical trial-and-error discovery unfeasible. This project leverages the Redtail AI Factory to computationally discover the next generation of PFAS-destroying nanocrystals within a 6-month timeframe. We will utilize a novel, AI-ready dataset comprising extensive experimental and quantum-chemical parameters of PFAS defluorination over diverse metal catalysts. By training a message-passing Graph Neural Network (GNN), we will extract the crucial latent factors governing C–F bond cleavage (e.g., optimal Sabatier-adjusted d-band centers, lattice strain, and coordination geometry). Using Transfer Learning, we will map these learned representations onto a massive secondary dataset of over 380,000 highly stable, previously untested inorganic crystals derived from the recent Google GNoME materials database. The computational power of Redtail will enable the ultrafast, high-throughput screening of these materials to identify the top 5 theoretical candidates for complete PFAS mineralization. The primary outcome will be an open-source predictive model and a curated catalyst candidate list, providing vital preliminary data for the PI's NSF (CAREER) proposal preparation and a potential proposal on advanced materials design for DOE AI Genesis Mission, and offering a transformative solution to Utah's PFAS issue in groundwater.

We propose two complementary efforts based in the department of Atmospheric Sciences that will use Redtail resources to transform atmospheric prediction for weather and climate AI. The first is PlumeCast, a topography-aware generative AI model that will reduce the expensive atmospheric particle trajectory simulations from weeks on a supercomputer to seconds on a single GPU and will be runnable across any landscape on Earth. Particle Dispersion Models are the gold standard for simulating transport of wildfire smoke, dust, and greenhouse gases, but a single 5-day, 200-trajectory run takes ~45 minutes on a CPU with larger applications requiring millions of runs. Using a diffusion model architecture with a transformer backbone, PlumeCast will reproduce the full turbulent ensemble in seconds and unlock real-time smoke forecasting, prescribed burn tools, and probabilistic dust forecasts with uncertainty quantification. The second is the Utah Snow Ensemble, an operational system that predicts snowfall over the western United States at lead times up to 10 days, combining NWP with AI-based statistical downscaling at 800-m grid spacing (far finer than the 18-25 km operational forecasts typically available at these lead times). We will use Redtail to optimize this snow ensemble, add ECMWF's generative AI forecasts to the ensemble, and upgrade snow-density prediction. Such efforts will lead to a prototype snowfall system for the 2034 Olympic Winter Games and help train University of Utah students on a hybrid NWP/AIWP system. Both efforts have AI-ready data on existing CHPC resources and are ready to use on Redtail immediately.

Publicly available brain tumor MRI remains underused for supervised segmentation because expert 3D annotations are costly, creating dependence on a few curated benchmarks and limiting confidence in cross-institutional generalization. We propose a six-month project to develop an anatomy-informed, topology-aware foundation model for unimodal glioma segmentation. Our premise is that glioma both disrupts normal neuroanatomy and introduces persistent pathological structure. We will create and pretrain on a large textbook of healthy brains using dense whole-brain anatomical supervision, then fine-tune for glioma segmentation with scalar field topology objectives based on persistence simplification, contour trees, and branch classification. The project is positioned for rapid execution. We have assembled and quality controlled a 4-TB corpus of 4,196 glioma patients from 15 heterogeneous datasets, developed the required open-source topological software, established the model design and preprocessing workflow, and identified additional cohorts for model-assisted annotation and neuroradiologist review. Redtail's multi-node H200 capacity, CPU resources, and large scratch storage, are essential for training-shard generation, distributed training, and evaluation across diverse cohorts. Targeted AI Engineer and CHPC support will establish a reproducible environment and address monitoring and scaling bottlenecks. Expected outcomes are a validated, openly released foundation model and reproducible workflows; neuroradiologist-audited annotations released through TCIA; a manuscript submission; and a federal grant proposal. By expanding reusable annotated data and improving robust glioma segmentation, this project will strengthen open medical-imaging AI research capacity and lay groundwork for broader clinical validation in Utah and beyond.

Quantitative measurements derived from medical images are essential for diagnosis, longitudinal assessment, and surgical planning, yet most remain manually performed. Existing imaging AI systems generally automate individual measurements using task-specific models, limiting their adaptability and scalability. We propose to train a vision-language model (VLM) to perform comprehensive quantitative reasoning from full-length scoliosis radiographs. The project is immediately ready for large-scale training. We have curated and anonymized 24,434 scoliosis radiographic studies from 13,786 subjects with paired radiology reports for multimodal pretraining; assembled 2,000 multi-institutional radiographs with machine-readable expert landmark annotations for supervised training; and reserved greater than 1,000 additional radiographs for held-out testing. The principal remaining barrier is computational scale. Using Redtail, we will perform multimodal domain adaptation followed by spatially supervised training and test whether image-report pretraining improves anatomical localization and quantitative spatial reasoning. The resulting model will derive comprehensive spinal alignment measurements from predicted anatomical landmarks and be tested across institutions and challenging postoperative and severe-deformity cases. This work will produce a quantitative spine VLM and establish a scalable framework for quantitative medical imaging beyond scoliosis.

This project aims to bridge the mixed-phase particle-to-cloud scale gap, a fundamental barrier to precipitation prediction critical for water security. Nanometer-scale aerosol chemistry controls ice nucleating particle (INP) concentrations that dictate kilometer-scale cloud glaciation via non-linear dynamics governing phase partitioning, cloud lifetime, and precipitation. Current models fail because empirical INP parameterizations ignore vertical structure, particle composition, altitudes, and atmospheric turbulence. BRIDGE's vision harnesses ARM/EMSL's vertically resolved, size-resolved aerosol composition and ice nucleation profiles to deliver a hierarchical AI pipeline. This fuses field measurements with models for unprecedented cross-scale precipitation prediction.

Clinical flow cytometry is essential for diagnosing and monitoring hematologic malignancies, but interpretation remains dependent on expert manual review, disease-specific gating strategies, and local laboratory practice. These constraints limit reproducibility and make it difficult to fully use the information contained in large archives of high-dimensional single-cell clinical data. EventHorizon is a foundation-model framework for clinical flow cytometry that uses self-supervised learning to generate event-level and case-level representations from approximately 100,000 de-identified clinical cases. The proposed Redtail Challenge project asks a focused, high-impact question: can a Redtail-scale distributed training regime produce larger EventHorizon models that measurably improve representation quality and downstream diagnostic performance? This project is well suited to the Redtail Challenge because it is AI-ready, scientifically mature, and constrained primarily by access to large-scale compute and distributed-training engineering. The dataset, baseline model architecture, preprocessing pipeline, and evaluation workflows already exist. Redtail access, AI Engineer support, and CHPC facilitation would enable a rapid and rigorous closed-end scaling study rather than an open-ended exploratory effort. The near-term milestone is to determine whether larger EventHorizon models justify continued development as clinical flow cytometry foundation models.

Today's network tools, such as a network Looking Glass, provide real-time expert information to network engineers that have extensive knowledge in running routers, switches, and firewalls. However, users and administrators without these skills need something to query in natural language that can give them hints and relevant information about their upstream network. By creating a chat agent that queries a Looking Glass like an experienced engineer, this project aims to provide K-12, library, university, and community anchor institutions with straightforward and understandable feedback about the statewide network around them. Building this agent to be responsive and reliable in a variety of user-created scenarios will require training and extensive validation. This pilot will explore a scalable solution with network engineers at the university level and administrators of varying skill levels at the school district/library system level throughout the state, with the eventual goal to lower skill barriers of exploring networks and identifying potential issues for any user tasked with technical troubleshooting.

Biomolecular binding controls nearly every cellular process and mediates the actions of most drugs, yet discovering these interactions remains slow and laborious. AI-driven structure prediction tools such as AlphaFold have the potential to accelerate this discovery. To take advantage of this technology, the Bosch Lab developed software called MOBILS that automates in silico binding predictions at the CHPC, enabling large-scale discovery screens. Over two years, we have performed screens for 12 Utah research groups studying topics including hormones, cancer, and virology. Despite this progress, in silico screens still suffer from questionable accuracy and long runtimes. Here, we propose to use Redtail to address these issues by benchmarking in silico binding screens. In one objective, we will generate predictions for 2M randomly paired negative controls and combine them with existing predictions for positive controls. This dataset will enable rigorous comparison of binding confidence scores at a scale not achieved before. In a parallel objective, we will test whether GPU-accelerated multiple sequence alignment improves screen speed and throughput. Redtail uniquely enables this work within 6 months, due to its large H200 GPU capacity and reconfigurable nodes. The first objective is ready to begin, while the second requires limited facilitation. Outcomes include evidence-based recommendations for selecting binding confidence scores, validated CHPC workflows, and dissemination of our findings on bioRxiv, GitHub, CHPC documentation webpages, and at local presentations. Together, these will enable faster, more reliable AI-driven biomolecular discovery across Utah and beyond.

High-throughput drug screening (HTS) is increasingly used in precision oncology, but its analysis often relies on ad hoc, spreadsheet-based workflows, limiting standardization and reproducibility. We previously developed DS5, an open-source Python framework that integrates storage, quality control, dose-response analysis, and cohort drug prioritization (gitlab.com/qiao-lab/ds5; manuscript under review). We now propose agentic-DS5, a natural-language extension that runs on self-hosted open-weight models, generates and executes custom analysis workflows, and retrieves and evaluates published evidence to support drug prioritization. Our proof-of-concept prototype, demonstrated to clinical and laboratory collaborators, uses Claude Sonnet 4.6 to translate natural-language requests into DS5 function calls. We will extend it in three directions. First, we will deploy and benchmark self-hosted open-weight models to determine whether they are sufficient for this workload and where they fall short, establishing an architecture that can eventually operate within an approved protected environment while avoiding per-query API charges. Second, we will enable custom analysis code generation, executed in safe, reproducible, sandboxed pipelines. Third, we will integrate literature retrieval for prioritized drugs, assessing evidence quality against predefined criteria and providing verifiable, claim-level citations. The result is an open, reproducible platform deployable by any laboratory with institutional computing access, enabled by Redtail capabilities beyond the reach of an individual laboratory: large-model serving, benchmarking capacity, and AI-engineering support. Through existing collaborations, initial adoption will be at precision oncology programs at Huntsman Cancer Institute and Intermountain Primary Children's Hospital, serving patients who have exhausted standard-of-care options and whose clinical window is measured in weeks.

Cryo-electron microscopy (cryo-EM) is a primary route towards obtaining high-resolution reconstructions of biomolecular targets that are central to drug discovery and life science research. However, routinely achieving high-resolution reconstructions remains limited by the phenomenon of "preferred orientation." This occurs when the 2D projections of the biomolecule from collected micrographs lack sufficient orientations to inform a high-resolution 3D reconstruction. Current experimental solutions to preferred orientation are costly and time-consuming, while existing computational solutions only utilize high-amplitude signals during reconstruction. This project reframes preferred orientation as a masked-imputation problem, where a generative model will first learn physically plausible 3D reconstructions from deposited high-resolution structures, and then be evaluated on its ability to identify experimentally derived low-amplitude signals containing information on missing poses. The generative imputation of missing orientations has not been demonstrated in cryo-EM to our knowledge, but has the potential to combat preferred orientation in ways existing methods cannot. Over six months, we will complete architecture selection, quantitative imputation benchmarking, and validation on real preferred-orientation-limited datasets, finally concluding with an explicit go/no-go recommendation. This project is a direct test of whether the University of Utah's AI Factory can turn frontier GPU infrastructure into a load-bearing tool for solving an open problem in biology. A validated result would give Utah's academic and industrial biotech and pharmaceutical sectors faster access to high-resolution structures while opening a fundamentally new route to investigate rare biomolecular states, thereby turning a limitation the field has tolerated for decades into a solvable data problem.

Advances in portable computing and digital image processing have accelerated the adoption of point-of-use, colorimetric biosensors and point-of-care biomedical diagnostics. However, their real-world utility remains restricted by spatial-chromatic cross-talk—where a target's perceived color shifts due to background colors, ambient lighting, display gamuts, and user perceptual variations. Digital cameras and image processing workflows interpret color signals differently from human eyes because of such cross-talk. Existing AI image classifiers use standard deep neural networks that discard granular color data early in processing layers, and so cannot resolve this cross-talk. Furthermore, standard computerized color correction relies on static calibration references that ignore these dynamic spatial-chromatic cross-talk interactions. This project proposes a novel Artificial Intelligence (AI) framework to model, predict, and computationally isolate cross-talk effects. By constraining neural networks with biophysics-informed color science, we will create an algorithmic normalization layer that standardizes biomarker quantification from digital images captured across variable testing environments, display gamuts, and users. We will validate this framework by quantifying reliability and reproducibility improvements in color images representative of colorimetric biosensor and biomedical diagnostic images.

This project proposes scaling Horizons, an AI-enabled curriculum development and review platform, by migrating it to the University of Utah's Redtail computing environment. The work will use open-weight language models, multi-agent workflows, and automated evaluation to improve curriculum design, accessibility, quality assurance, and program review at scale. Expected outcomes include fine-tuned educational models, reproducible evaluation workflows, and a secure, sustainable institutional framework for using AI responsibly in program redesign while preserving faculty oversight and accountability.

Interfacial thermal transport governs the performance and reliability of next-generation power electronics, radio-frequency devices, and thermal management systems built on wide-bandgap and ultra-wide-bandgap semiconductor heterostructures (GaN, AlN, diamond, Si, and metal contacts). Predicting thermal boundary conductance across these interfaces at the accuracy needed for design requires interatomic potentials that combine first-principles fidelity with the scale of classical molecular dynamics, a combination now achievable through machine-learned interatomic potentials (MLIPs), but only if MLIP training, molecular dynamics production runs, and phonon Boltzmann transport equation (BTE) calculations can be run at far greater scale and speed than conventional CHPC allocations allow. This project will use Redtail's AI Factory infrastructure to accelerate an active pipeline for interfacial thermal transport prediction: (1) scaling AIMD dataset generation and MLIP training (MTP, DeepMD) across multiple heterostructure systems (GaN/metals, Si/diamond, GaN/diamond, GaN/AlN, and disordered Al1-xGaxN alloys), (2) accelerating an in-house, end-to-end differentiable JAX framework for fitting interatomic potentials, and (3) running large-batch nonequilibrium MD (NEMD) and BTE transport calculations across a systematic sweep of interface chemistries and disorder levels. The outcomes directly inform thermal management strategies for power electronics relevant to Utah's semiconductor base, and will be disseminated through peer-reviewed publication, an open-source MLIP/data release, and use as preliminary results in a follow-on federal proposal (e.g., NSF).

Utah institutions of higher education must comply with federal professional licensure disclosure requirements under 34 C.F.R. § 668.43 by informing students whether academic programs meet licensure requirements in the states where students reside. This obligation is increasingly complex because online learners are located nationwide and licensure requirements vary across states, professions, and regulatory structures. Through its role as Utah's State Authorization Reciprocity Agreement (SARA) portal agency, the Utah System of Higher Education (USHE) supports a collaborative state-by-state professional licensure resource used by 15 Utah institutions. Maintaining this resource requires extensive manual review of thousands of statutes, regulations, licensing board websites, and other official sources. This project will evaluate whether artificial intelligence can reliably support and scale the professional licensure research process while preserving transparency, source traceability, and human oversight. Using a validated benchmark dataset developed through extensive nursing licensure research in Utah, Idaho, and Alaska, the project will focus on four nursing professions: Registered Nurse (RN), Licensed Practical Nurse (LPN), Advanced Practice Registered Nurse (APRN), and Certified Nurse Aide (CNA). Leveraging the University of Utah's Redtail high-performance computing environment, the team will test AI methods for source identification, retrieval, structured data extraction, citation preservation, requirement comparison, and human-review triage. Project outcomes include a human-validated nursing benchmark and data dictionary, a Redtail-based research pipeline, a comprehensive evaluation of AI accuracy and limitations, and an expanded nursing licensure dataset covering all fifty states. Long term, the project aims to reduce duplicative compliance work across Utah institutions, improve licensure disclosures and student advising, and create a scalable, evidence-based research framework that can be extended to additional licensed professions, ultimately benefiting Utah students seeking employment nationwide.

Tuberculosis (TB) remains a major global health burden, and the immune mechanisms that allow some infected individuals to control Mycobacterium tuberculosis (Mtb) while others develop active disease remain incompletely understood. Growing evidence indicates that antibodies contribute to protective anti-TB immunity, yet a central bottleneck is determining which Mtb antigens are recognized by disease-associated antibody sequences. Experimental screening of thousands of antibodies against large antigen panels is slow and costly. We will use Redtail to test whether state-of-the-art protein structure models can provide a scalable first-pass method for ranking antibody–antigen interactions. Our AI-ready dataset contains approximately 20,000 paired heavy- and light-chain B-cell receptor sequences generated by single-cell sequencing of Mtb antigen-enriched memory B cells from four individuals with latent TB infection and four with active TB. We have already implemented a pilot workflow that models five monoclonal antibodies against ESAT-6, CFP-10, and Ag85A using AlphaFold 3 and Boltz, multiple random seeds, monomeric and multimeric antigen configurations, and complementary confidence and physical-interface metrics. During this six-month project, we will benchmark model performance using experimentally characterized and decoy antibody–antigen pairs, establish a reproducible composite ranking score, and scale the optimized workflow to representative latent- and active-TB-associated clonotypes. The project will produce a benchmarked computational pipeline, a ranked set of candidate antibody-antigen interactions for experimental follow-up, and an initial map of antigenic targets that distinguish controlled from active TB. These outputs could accelerate discovery of protective antibodies, vaccine antigens, and repertoire-based diagnostics.

To be ready for the careers ahead, students must learn to build with AI, not just chat with it. Building with AI means writing software that calls a model through an API (application programming interface), the way every real AI product is made. And API access is exactly where student learning breaks down today. Commercial AI APIs require credit cards and bill per request, so instructors cannot assign real AI projects to a class, hackathon organizers cannot promise every team a key, and student founders prototype on personal money. We propose a six-month pilot of the Utah Student AI Gateway: a shared AI service for University of Utah students, hosted on the Redtail AI Factory. The service will run capable open-weight models (DeepSeek V4 Flash, Qwen3.8 27B, NVIDIA Nemotron, and Meta Muse Glimmer) on Redtail GPUs, swapping in newer models over time. Students sign in with University accounts, receive a personal API key with a monthly usage limit, and build: first vibe-coded apps, then their own AI-powered products and startups. The pilot centers on student entrepreneurs: the Computer Science Sandbox course, where students build software startups, and Lassonde Entrepreneur Institute students, to whom Lassonde will offer the Gateway as an option. The pilot is deliberately bounded: a 500-student cap, metered usage, a public dashboard, and a fixed end date. It ends with a public report and an open-source deployment guide any Utah System of Higher Education institution can reuse. Utah students graduate having built with AI, on Utah's own infrastructure.

BIRDBASE is a global avian trait dataset covering 78 published traits, represented by 97 auditable biological fields, for all 11,589 recognized bird species. Built over 26 years and published in Scientific Data in 2025, it is ready for a high-impact AI completion sprint. Remaining blank cells include breeding, nesting, movement, elevation, morphology, behavior, habitat, and other traits for rare and poorly studied species. We will use Redtail to evaluate every eligible blank cell against Birds of the World and accessible ornithological literature. A pipeline will parse and index authorized sources, retrieve species- and field-specific passages, serve open-weight language models on H200 GPUs, and return candidate additions with exact evidence and provenance. A second local verification pass tests whether each citation supports the proposed value. Our acceptance-tested v27.2.2 rules engine rejects overwrites, unsupported inference, invalid ranges or units, taxonomy changes, legacy-mass revisions, and unsafe proposals; unsupported cells remain blank. The project is already operational: a 30-species pilot yielded 20 source-supported additions while withholding unsupported values, and the validator passes 58 of 58 synthetic acceptance checks. Graduate and undergraduate lab members will support adjudication, quality control, and production. Redtail access, requested through November 30, 2026, will let us complete a BIRDBASE 2.0 release candidate, cell-level provenance and audit files, and a reproducible workflow. The team will then submit the accompanying data paper to the journal Ecology by December 31, 2026, converting Redtail's scale into a durable, open biodiversity resource and a reusable model for auditable extraction of any type of biodiversity data.

The MAGIC-SCAN/FASTMAP project (https://www.price.utah.edu/2024/08/13/the-u-collaborates-with-tulane-on-up-to-23-million-cancer-moonshot-project-to-build-advanced-tumor-imaging-system) tackles a longstanding challenge in surgical oncology: the inability to verify complete tumor removal in real time. Positive surgical margins occur in 30% of surgical procedures and are typically detected a week or more post-operation. This delay forces patients and care teams to make difficult choices between venturing into risky follow-up surgeries or simply waiting and monitoring for potential recurrence. To empower surgeons with immediate procedural success, this project unites the world's fastest high-resolution scanner (MAGIC-SCAN developed at Tulane) with an advanced cyber-infrastructure (FASTMAP developed by the CEDMAV team in Utah) for automated intraoperative cancer detection. By utilizing the Redtail AI Factory, we will train state-of-the-art AI models on massive medical imaging datasets. We aim to deliver 0.5 μm-resolution virtual pathology results within a maximum of 10 minutes, meeting a benchmark set by the ARPA-H Precision Surgical Interventions program. This target includes in situ staining and scanning of the dissected organ (5–6 minutes), as well as AI inference and visualization (4–5 minutes) for the surgeon. MAGIC-SCAN employs optical sectioning super-resolution structured illumination microscopy (OS-SR-SIM) to acquire terapixel-scale data, generating up to 2 TB of data per patient for real-time AI inference. Successfully benchmarking and optimization of this data volume demands the high computational density of the Redtail system to convert exploratory research into an operational clinical solution. Redtail will be used to develop hundreds-of-terabyte data fabrics for training and real-time terabyte data inference at the edge to support the human benefit of detecting elusive residual cancer in the operating room.