Microscales
Explore Ultra

Microscales. Measurement-aware models for science, and Ultra, the workspace where researchers use them.

Microscales

Measurement-aware models for science.

Scroll to explore

We are building models that understand scientific measurements, starting with microscopy, and Ultra, the workspace where researchers use them to execute and inspect analyses.

Our first evidence is in microscopy. The long-term goal is a model family spanning signals, images, volumes, channels and time.

Science model

Access to an instrument file is not the same as understanding its measurements.

A valid analysis keeps physical scale, channel identity, acquisition metadata, spatial structure and change over time intact. We are training models that read measurements with those properties, and we say plainly what they can do today.

What Ultra supports today and what the learned model has been evaluated on, by data dimension
DataIn the Ultra workspaceLearned model
1DSignals and spectraToday Tables, plots and code on Python workersRoadmap Not yet evaluated
2DMicrographs and detector framesToday Imaging viewers, analysis, provenanceEvaluated Paired-microscopy structure identification, September 2026 checkpoint
3DVolumesToday OME-NGFF volume viewersRoadmap Not yet evaluated
4DTime-lapse volumesToday OME-NGFF datasets, with the axes stored in the fileRoadmap Not yet evaluated
5DMulti-channel time-lapse volumes (x, y, z, channel, time)Today The OME data model, opened as storedRoadmap Not yet evaluated
60.4%
Brightfield top-1
97.9%
Fluorescence top-1
4.2%
Chance baseline

Early evidence · September 2026 checkpoint

The task: name the structure in a paired microscope view the model had never seen, choosing among 24 candidates. Held out: 48 crops. Checkpoint step 12,800, internal audit.

What it does not show: general understanding across dimensions, segmentation quality, calibrated uncertainty, or downstream scientific conclusions. Chance is a sanity check, not the comparison that matters.

The next evaluation: the same task with a general agent and existing tools, the same agent plus this model, and the current expert workflow, on the same data with the same acceptance criteria.

General agents

Extending general agents with scientific measurement capabilities.

The model layer is being built as a tool a general agent can call. The question we want to answer is simple: does a specialized scientific model make a general agent materially better at instrument-data analysis?

  1. 01

    General agent

    Claude or another model interprets the request and coordinates the task.

  2. 02

    Microscales model and scientific tools

    Perception and measurement on the source data, not on a screenshot of it.

  3. 03

    Structured outputs

    Structures, measurements, scale and uncertainty, tied to the source measurements.

  4. 04

    Review

    Presented for inspection in Ultra or another application, with the run recorded.

Ultra’s own agents already run on the model you configure: OpenAI-compatible endpoints, vLLM, Ollama, or Claude through Amazon Bedrock. The tool interface for the science model is in development. No partnership with any model provider is implied.

Ultra

Explore, run, and inspect. In one place.

Ultra brings scientific viewers, computational tools and durable agent execution together, so a question can become a reproducible result without leaving the workspace.

Connected regions by minimum area
Regions · Otsu 38 · 512 × 512 px
Minimum area 1 px293
Minimum area 32 px285
Minimum area 64 px (default)248
Minimum area 128 px58
Measured by this website’s own analysis code on a public image: human cells undergoing mitosis (David Root, CC0; Moffat et al., Cell 2006), 512 × 512 px, lossless from the source TIFF (SHA-256 2751ba66…08cb456). Otsu threshold 38, 8-connected components. Bright connected regions, not a validated cell count. Not a benchmark of the hosted Ultra service.
Connected regions by minimum area, Otsu threshold 38
Minimum areaRegions
1 px293
32 px285
64 px (default)248
128 px58
Foreground pixels retained
Percent of 262,144 pixels
Minimum area 1 px11.4%
Minimum area 32 px11.3%
Minimum area 64 px (default)10.6%
Minimum area 128 px4.0%
Pixels above the threshold that survive the minimum-area filter: 29,781 · 29,716 · 27,667 · 10,570 of 262,144. Same image, same threshold, same independently verified run as the regions chart.
Foreground pixels retained by minimum area
Minimum areaPixelsPercent
1 px29,78111.4%
32 px29,71611.3%
64 px (default)27,66710.6%
128 px10,5704.0%
Browser result against an independent reference
Parameter sets in agreement
Threshold and region counts100%
Foreground pixel counts100%
Full masks, SHA-256100%
Region centroids100%
The browser algorithm is checked against an independent NumPy and breadth-first-search reference (tools/verify-analysis.py, tools/check-analysis.mjs): five parameter sets from 1 px to 4,096 px minimum area, full masks and centroids, plus four edge cases. Every check passes.
Cross-check agreement, browser versus reference
CheckParameter sets matching
Threshold and region counts5 of 5
Foreground pixel counts5 of 5
Full masks, SHA-2565 of 5
Region centroids5 of 5

Drag to explore

Workspace

Every scale of your data, side by side.

Images, volumes, tables and notes live together. Viewers for scientific imaging and OME-NGFF volumes sit next to the analyses that read them. The walkthrough below is a product mockup; the live workspace is at ultra.ece.ucsb.edu.

ultra · workspace / cell-atlas-09
plate-04_dapi.tif · 2,048 × 2,048 · 16-bitsegment nuclei
organoid_stack.zarr · OME-NGFF · 48 slicesz 12 / 48
decay_fit.ipynb · Python workerrunning
fit = ultra.fit(dataset="plate-04", model="exp-decay")
fit.plot(provenance=True)

00:00|00:09

Auto-playing views

Product mockup, not a recording: open a plate, scan the field, and let Ultra segment nuclei with a model you choose.

Product mockup, not a recording: step through an OME-NGFF volume slice by slice, with axes and scale carried from the file.

Product mockup, not a recording: fit a model in a notebook cell backed by a Python worker, with the plot rendered as it runs.

Product mockup, not a recording: delegate a multi-step analysis to an agent; every step records what it used and produced.

Plate II. A transmission electron microscope column drawn as a patent figure, beside a simulated cryo-EM micrograph and the atomic model of the AAV2 capsid from PDB entry 6E9D.
Plate II · Transmission electron microscopy, single-particle analysis Capsid: adeno-associated virus 2 (L336C variant), PDB 6E9D, 1.86 Å cryo-EM, Tan et al. 2018. Micrograph simulated from the model.

Durable execution

From question to result, with the receipts.

  1. 01

    Explore

    Open images, volumes and tables together. Ask in plain language; Ultra pulls the right context into view.

  2. 02

    Execute

    Analyses run as durable jobs on Python workers, with scientific tools the agent can call. Close the tab and the work keeps going, then reports back.

  3. 03

    Inspect

    Every result keeps its inputs, code and parameters. Reproduce it, compare it, and share it with confidence.

Provenance

Every result keeps its lineage.

Step through a reference dataset and the result Ultra produced from it. Each slide of the output points back to the inputs it was made from.

Reference data

Micrographplaceholder
Metadataplaceholder
Raw signalplaceholder
Field notesplaceholder
1 / 4input

Ultra result

Segmentationplaceholder
Provenance graphplaceholder
Fitted modelplaceholder
Reportplaceholder
1 / 4output

Ultra, built for research

Serious tools, quietly arranged.

Ultra combines a web workspace, scientific image and volume viewers, durable agent execution and computational tools on one control plane.

Scientific viewers

Imaging and OME-NGFF volume viewers open multi-gigabyte data in the browser, with metadata and scale carried from the file.

Durable agent execution

Long-running analyses execute on Python workers coordinated by a Go control plane. They survive closed tabs and report back when done.

Provenance by default

Results are stored with the inputs, code and parameters that produced them, so anything in the workspace can be traced and rerun.

90+
Scientific file formats decoded by the imaging service
4,600
Passing tests across the control plane, agents and viewers
24/7
Durable runs that outlive the browser session
1
Workspace for images, volumes, code and notes

Format count from the imaging service’s documentation; test count from the Ultra repository’s September 2026 validation record (2,288 Python and 2,313 frontend tests passing).

Ultra by Microscales

Meet Ultra.

Ultra is where the models meet the scientist: explore data, run analyses, and inspect every result with its lineage.

Explore Ultra Design partnerships