Preserve the source.
Native captures should remain available beside enhanced images and model outputs, so reviewers can inspect how an interpretation was produced.
Optics and intelligence.
Engineered together.
Throughput depends on the whole path: illumination, sensing, timing, reconstruction, inference and review.
01 / PHOTONSOptics
Illumination, contrast and calibration
02 / SIGNALSAcquisition
Sensors, triggers and synchronisation
03 / IMAGESComputation
Reconstruction and image processing
04 / CONTEXTInference
Local models and structured outputs
05 / EVIDENCEReview
Native data, annotations and QA
Measure the bottleneck.
Capture rate, reconstruction time, inference latency and data movement are separate quantities. Performance claims should name the workload and hardware.
Compute locally.
Spum runs supported models on local hardware. A complete instrument system also needs acquisition interfaces, application logic and validation.
Spum / Native inference
The engine inside the workflow.
Language models, image understanding and AfterGlow piano continuation run on the native engine. The Spum site documents supported hardware and benchmark conditions.
Explore measurements ↗ExecutionLocal hardware
InputsText · images · MIDI
Application focusBiomedical instruments
Architecture principleKeep inference close to capture
Build with Neuroscape
Let’s find the next step ↗