Ask a question about a microscope image.
VeriCell finds every cell or organoid, checks each one, and tells you how sure it is, plus which few objects a human should double-check.
1Pick an image2Ask (type or speak)3Watch it check its own work
1 · Image
2 · Question
3 · Reasoning model
Pick an image and ask a question to start.
outline colour = how sure VeriCell is it's real:unsuresure· gold = the object the answer is about · blue = the agent asks you to check
What the agent is doing
Each step and every decision the agent makes will appear here, in plain words.
Technical trace
How VeriCell works
A fixed pipeline does the heavy work, Muse Spark makes decisions only where judgement is needed, and plain code computes every number, so the AI can't invent one.
1 · Find
- Muse Spark reads the image and plans the question
- Cellpose-SAM and SAM 3.1 find every object, zoomed in
- A verifier scores each object: how likely is it real?
2 · Check (the agent)
- Muse Spark chooses one action at a time: judge a crop, re-ask SAM, or ask you
- A safety gate blocks invalid actions and enforces budgets
- After each action the answer is re-measured; it stops when sure
3 · Answer
- Exact geometry, no AI arithmetic
- A 90% range, calibrated on held-out images
- Your checks narrow the range live
Does it work?
Organoid benchmark
16 / 16
question types with lower error than the published MorphoOrgaAgent
Largest-organoid area error
3.6%
vs 16.8% published
Spots its own wrong cells
0.866
AUROC, vs 0.823 for SAM's own confidence
Honest confidence
0.012
calibration error: "80% sure" means right 80% of the time
90% ranges contain the truth
89–92%
on images it never trained on
Open weights
16 / 16
plans correct on Muse Spark Open (Fireworks), same as hosted
Cell
Click a cell, or pick one from the review queue.