In May 2026, a demo video of an AI-built interactive 3D cell explorer racked up 480,000 views on X and earned nearly 6,000 likes—not from a funded edtech startup, but from a weekend project by software engineer Dilum Sanjaya (@DilumSanjaya).
The reaction from educators was immediate. Jesse Genet (@jessegenet), whose X bio notes she's homeschooling four kids, wrote: "Education is going to get so fun y'all 🤯" A homeschooling dad in the replies called it a game-changer for teaching biology at home.
This is the third app in Sanjaya's "interactive science app ideas" series—following a solar system flyover and a black hole visualization. Each was built using AI tools, not months of traditional development. Here's how the cell explorer works, what made the viral response possible, and what the underlying tech stack signals about where AI-assisted development is headed.
What the 3D Cell Architecture Studio actually does
The Cell Architecture Studio is an interactive 3D application that lets users:
- Rotate and zoom 3D biological structures in real time
- Dissect cells to expose interior organelles
- Browse side panels with annotated facts about each structure
- Switch between microscope-style views and full 3D perspectives
- Compare cells side-by-side (plant cell vs. neuron, for example)
The modeled structures include plant cells, neurons, and other biological architectures. The depth of each model—with named and labeled organelles, color coding, and interactive layers—is what caught the attention of educators. This isn't a static diagram with a hover tooltip; it's closer to a virtual biology lab.
The three-tool AI stack
Dilum explicitly credited three AI tools in his demo post:
GPT Images 2 — UI design
GPT Images 2 (OpenAI's current image generation model, API name gpt-image-2) handled the UI design. Rather than hiring a designer or spending days in Figma, Sanjaya used image generation to establish the visual aesthetic, layout structure, and interface elements.
This is an underappreciated use case for image generation models. Most public attention focuses on art generation and product imagery, but using models like GPT Images 2 to generate UI mockups that a developer then implements is increasingly common among solo builders. Our explainer on GPT Images 2 covers the model's resolution options and API access details.
Gemini 3.1 Pro — code generation
Gemini 3.1 Pro (Google DeepMind) handled the application code. For an interactive 3D app with user-controlled camera movements, data panels, and smooth state transitions, this is non-trivial—a project that would have taken a solo developer weeks to scaffold from scratch.
This is vibe coding in its most literal sense: the developer describes behavior and constraints in natural language, and the frontier model produces working code. The developer's role shifts from typing implementation to directing and verifying it.
Tripo — 3D model generation
Tripo is a 3D AI generation platform that converts text or image prompts into 3D geometry. For the cell explorer, Tripo generated the actual 3D meshes of biological structures—the models that users manipulate in the app.
This is the piece that would have been most prohibitive before AI 3D tools existed. Producing accurate, visually appealing 3D models of cells and organelles would have required specialist 3D artists or hours of manual modeling in Blender or similar tools. Tripo makes it a prompt.
Why the combination worked
The insight in Sanjaya's stack is division of labor across specialized AI tools:
| Layer | Tool | What it replaced |
|---|---|---|
| Visual design | GPT Images 2 | UI/UX designer + mockup workflow |
| Application logic | Gemini 3.1 Pro | Frontend + 3D interaction engineer |
| 3D assets | Tripo | 3D modeler + hours of manual work |
None of these tools individually could have produced the app. But combined, they enabled one developer to ship something that would have needed a small cross-functional team six months earlier.
The meta-skill is knowing which specialized tool to apply to which layer of the problem—not just using one AI model for everything.
The education response
The viral reaction wasn't just developer admiration—it was educators recognizing a new category of learning tool.
Biology education has always struggled with the gap between static textbook diagrams and the dynamic 3D reality of cellular biology. Even the best biology textbooks reduce cells to flat, labeled cross-sections. Interactive tools that let students explore rather than just read have existed—but historically required expensive lab software licenses or custom institutional development.
What makes the AI-built approach different:
- Cost: One developer, a few AI tool subscriptions, and a few days of work—not a six-figure development contract
- Iteration speed: Feedback from educators can be turned into product changes in hours, not sprint cycles
- Extensibility: The same stack can generate new cell types, new organisms, and new educational levels without starting over
The homeschooling community responded especially strongly—because off-the-shelf curricula rarely include interactive 3D biology tools, and paying for institutional edtech software is impractical for most families.
The "data is the hard part" counterpoint
Not everyone in the thread was uncritical. Shae McLaughlin (@shae_mcl), a bioengineering PhD at Stanford and Arc Institute, offered a grounding observation:
"It's estimated that the Protein Data Bank (PDB) cost around $13B to create. AlphaFold was only possible because of it. If we want ML to solve biology, we should be funding the creation of databases and the development of new assay technologies. ML is nothing without data."
This is worth sitting with. The Cell Architecture Studio is a presentation layer—it makes existing biological knowledge more accessible and engaging. The hard work of generating biological knowledge (experimental data, measurements, wet lab validation) remains a human and institutional challenge that AI visualization tools don't address.
But the criticism and the celebration aren't mutually exclusive. We need both better biological databases and better tools for communicating what we already know. Sanjaya's app does the latter exceptionally well.
A template for AI-assisted science apps
The cell explorer suggests a repeatable template for solo developers building educational tools:
- Identify a domain where 3D or interactive visualization is currently poor — biology, chemistry, physics, anatomy, geology
- Generate 3D assets with tools like Tripo, Meshy, or similar platforms
- Build the interaction layer using a frontier code model (Gemini, Claude, GPT-4o)
- Design the interface with image generation for rapid UI prototyping
- Validate with a short demo video — feedback arrives in hours
The feedback loop from demo to viral to real users is now measured in hours, not quarters. For science education in particular, that speed matters: the teachers and parents who responded to Sanjaya's post are a direct feedback channel that most edtech companies would pay significantly to access.
What's next for Sanjaya's series
The series pattern suggests Part 4 is coming—and with each iteration, the stack likely gets more refined. The interesting constraint in the next app won't be the AI tools (those are improving fast) but which domain has the right combination of visual complexity, educational value, and existing 3D-able reference data.
Chemistry (molecular structures), anatomy (organ systems), and environmental science (ecosystems) all seem like natural candidates.
Related reading on ExplainX
- ChatGPT Images 2.0 and gpt-image-2: what builders should know
- How do diffusion image generation models work?
- What are agent skills and why they matter for AI development
- Stanford AI Index 2026: key takeaways for developers
Viral demos are not peer-reviewed evaluations. Tool capabilities, API availability, and pricing for GPT Images 2, Gemini 3.1 Pro, and Tripo can change; verify current specs on each provider's documentation before building production apps.