CozyBio
No-code 3D image analysis for wet-lab biologists. Node-based workflows, real segmentation.
What I Found
Every single lab counts cells by hand in ImageJ. One researcher told me she “counts by hand (hate it).” Labs acquire full 3D confocal stacks but collapse them to a single 2D plane for analysis, because no free tool handles volumetric segmentation reliably. The only 3D options are IMARIS and Volocity: $10K+ licenses, not scriptable, not reproducible. A typical pipeline touches 3 to 6 disconnected tools with manual export and reimport at every handoff.
What I Built (so far...)
I designed and shipped a working prototype in under a week: a browser-based, node-based workflow builder for 3D confocal stack imaging. Load a z-stack, run Cellpose 3D segmentation, measure per-object fluorescence intensity, export CSV. The full pipeline that replaces the most common manual ImageJ workflow.
The architecture uses OME-Zarr with zero-copy data references, a DAG-based execution engine with per-item caching and cooperative cancellation, and a Neuroglancer-based 3D viewer. I built it this fast by going deep on agentic AI development: I defined TDD workflows, custom skills, and cursor rules, then used Cursor and Claude Code to execute against my architecture and PRD.
Early Feedback
“If I could have a software to easily and accurately quantify cell-cell interactions, I will be able to merge these data with spatial-omics of the same imaged samples, leading to something magic. In the context of prediction of prognosis this will be of immense help.”
Build Log
Discovery
Interviewed 6+ researchers across multiple labs. Analyzed imaging workflows from six published papers spanning tumor organoids, vascularized tumor models, and cortical organoids. Every single lab counts cells by hand in ImageJ. Built a detailed PRD grounded in specific pain points, not assumptions.
Architecture + Build Sprint
Designed the full system architecture: node-based DAG execution engine, zero-copy OME-Zarr pipelines, Neuroglancer-based 3D viewer with a dual-layer trick for label overlays. Then used agentic AI development (Cursor, Claude Code, TDD, custom skills and rules) to ship a working prototype in under a week. Four build phases completed: layout and data pool, workflow engine, Cellpose 3D segmentation, and per-object measurement with CSV export.
Build, Measure, Learn
Researchers testing the prototype. Collecting feedback on segmentation reliability, workflow fit, and missing features. Visited the Analytica trade show in Munich, talked to Leica’s microscopy tool team about their approach. Their main insight matched mine: the problem isn’t applying segmentation, it’s reliability and accessibility.