Computer Vision in Little Rock, AR: Vision for the State Capital's Health, Defense, and Logistics Spine

Realistically, sixty to one hundred fifty days from first contact to signed DUA, depending on the type of data and whether you are accessing PHI, de-identified imaging, or a research warehouse extract. That is before any IRB review, which adds another thirty to ninety days for retrospective imaging studies. A CV project plan that promises a model in three months has not accounted for either of those steps. A capable Little Rock partner working with UAMS plans the modeling work to start in month four or five and uses the early months for protocol design, dataset specification, and clinical-collaborator alignment. Skipping the protocol work to save calendar time produces models that pass internal metrics and fail clinical validation.

Only in coordination with Dassault's quality and FAA airworthiness teams, and only if the partner is prepared for the documentation load. Anything that touches an aircraft inspection record is part of the certification trail, and that means the model itself, the training data lineage, and the change-control process for retraining all become artifacts the FAA and Dassault internal quality may want to see. Practical engagements are scoped around assist tools — vision systems that surface candidate findings to a human inspector who still owns the call — rather than autonomous inspection. The right Little Rock partner has either worked at Dassault, at a Tier 1 aerospace supplier, or at one of the FAA-regulated MROs in the region and treats the documentation effort as part of the project, not an afterthought.

Most production training runs on AWS, Azure, or sometimes Lambda for spot capacity, because that is where UAMS's research cloud agreement and most carriers' enterprise IT live. Local options are narrower. UAMS has internal HPC for sponsored research. UA Little Rock and the Emerging Analytics Center have GPU resources accessible through specific collaborations. A handful of local data-center operators offer co-location that can host a private GPU rack, which makes sense for buyers with steady inference loads but rarely for training. The honest answer is that compute is a mostly-solved problem in Little Rock; the bottleneck is usually data access and annotation, not GPU availability.

For a single terminal with two to four gate cameras and a dozen dock-door cameras, expect a first-year cost between one hundred ten and two hundred forty thousand dollars all-in. That includes the camera and pole hardware, edge compute, model training and tuning, integration to McLeod or whatever TMS the carrier runs, dashboard and alerting work, and a quarterly retraining cycle. Multi-terminal deployments scale efficiently — the second and third terminals come in significantly cheaper because the model and integration are reused. The number that surprises buyers is annual operating cost: figure twelve to twenty percent of the install cost per year, mostly for retraining and the on-call support relationship.

More than people realize. The state has a working pool of senior CV engineers across UAMS Radiology research, the Acxiom data-science alumni network, the Dassault and Caterpillar manufacturing-IT teams, the Welspun and LM Wind plant engineering rolls, and the Arkansas Children's Hospital research informatics group. Public surfaces include the UA Little Rock Emerging Analytics Center events on Capitol Avenue, the Venture Center's occasional AI nights downtown, the Arkansas Research Alliance speaker series, and meetups that float between the Argenta district and the River Market. A CV consultant who is plugged into two or three of those circles can credibly claim Little Rock-bench depth; one whose only Arkansas reference is a LinkedIn search probably cannot.