01NVIDIA Engineering Lab
We test the NVIDIA stack on real workloads before our clients bet on it.
Most teams can get a Blueprint running. Fewer can say what it costs per answer, which GPU it needs or why retrieval fails on scanned documents. That is what this lab measures, and it is how we deliver NVIDIA AI for clients.
01.1Inference and GPU cost
Serving benchmarks across model sizes, quantisation and batching, turned into cost per task.
NIMTensorRT-LLMDynamo
01.2Retrieval and guardrails
Embedding, re-ranking and guardrail placement evaluated on healthcare and lending documents.
NeMo RetrieverNeMo Guardrails
01.3Vision at the edge
Detection and video search on low-light CCTV and aerial feeds, sized for edge hardware.
MetropolisDeepStreamJetson
Spirometry & digital signal processing
Lung function measurement is a signal problem dressed up as a medical device. We work on the parts that decide whether a reading is trustworthy: flow-volume curve reconstruction, noise and artefact rejection, drift correction, and effort quality detection — so a clinician sees a result they can act on rather than a number that needs interpreting.
DSPFilteringEmbeddedMedical device
03Diabetic retinopathy
Screening is the bottleneck, not diagnosis. Our work targets retinal image quality assessment, lesion detection and grading — with an obsession about the cases the model should refuse to call. A screening system that knows when to escalate to a human is worth more than one with a better headline accuracy number.
Medical imagingClassificationUncertaintyScreening
04AI pipelines and models fitted to a specific business case
Most enterprise AI fails on architecture, not intelligence. This track is about matching the model to the job: when a small fine-tuned model beats a frontier one, where retrieval belongs, how evaluation is designed before a line of code is written, and what the pipeline costs to run at volume on day 400 rather than day 4.
Model selectionRAGEvaluationCost at scale
05Soil moisture sensing & rice growth
Rice is grown on intuition and irrigated on habit. We work on soil moisture estimation from low-cost sensing, calibration against real field conditions, and correlating moisture profiles with growth stage — the groundwork for telling a farmer when to irrigate instead of telling them what a dashboard says.
SensingCalibrationAgri-techField data
06Experiments in vision
An open track, and deliberately so. Detection and tracking under conditions the papers never test — bad light, motion blur, cheap optics, aerial angles from drone work. Video analytics that runs at the edge instead of shipping every frame to a GPU somewhere. Some of this graduates into VuFindr. Some of it stays an experiment, which is the point of having a track for it.
Object detectionTrackingEdge inferenceAerial imageryFeeds VuFindr
07OCR and semantic models for healthcare documents
Healthcare still moves on paper that has been printed, annotated by hand, faxed and scanned before any system sees it. The easy part is reading the characters. The hard part is reading the document: which box on a claim form that number belongs to, which table row survived the page break, which pen mark overrides the printed value.
So the track has two halves. Layout-aware extraction on genuinely degraded input — deskew, low-DPI faxes, stamps and handwriting over print, tables that split across pages. Then a semantic layer that maps what came off the page to the concepts a payer or provider system actually consumes: member, plan, provider identifier, service line, date of service, denial reason. Character accuracy is not the metric. Whether the field was right is.
The other half of being right is knowing when you are not. We spend as much time on the confidence model — which fields the system is allowed to fill on its own and which get routed to a person — as on extraction itself. This is the research line behind DocuFindr.
OCRDocument layout analysisHandwritingSemantic modelsField-level accuracyHuman-in-the-loopFeeds DocuFindr
10decoders
