Inspiration
I was inspired by the movie Spiderman and spiders on Nat geo wild
What it does
Web design creates configurable webs of your idea.
How we built it
I built it using React, TypeScript, and React Flow (xyflow/react) for the canvas itself. Every stage of the project is a column, and every option is a draggable node, and picking one dims its siblings and rewires the animated connection forward, the same way a spider reinforces one strand of a web and lets the others go slack. A small Node.js server sits behind the frontend and talks to two different AI providers: Google Gemini for cloud generation, and a local Ollama instance for on device generation plus fast background tasks like scoring the cost/latency/strain of a chosen path and classifying whether two connected decisions support, require, or conflict with each other. Projects autosave to localStorage so you can leave and resume a web in progress, and everything can be exported as either a human readable brief or a structured JSON "build specification" designed to be handed straight to another AI assistant to implement. Integration with Backboard gives the project persistent decision memory logging every choice as a timestamped event so the reasoning behind a web doesn't get lost.
Challenges we ran into
The biggest challenge was training the AI, earlier on every generated project converged on the same handful of generic stage names "Data Source," "Core Logic," "Storage Layer" no matter what idea you typed in, which made every web look identical. I Fixed that by writing much stricter prompts and a validation layer that rejects vague, interchangeable-sounding options and forces the model to name real, specific technologies instead. I also had to think carefully about trusting the estimates a model returns (cost, latency, architectural strain) they are just planning guesses, not measurements, so the UI and the exported JSON both say that explicitly rather than let a confident-looking number pass as fact.
Accomplishments that we're proud of
Getting the web to be genuinely dynamic — three stages or seven, two options or four, entirely driven by what the AI (or you) put into it, with zero hardcoded layout. We're also proud of the handoff JSON: it's not just a data dump, it's a self-contained build spec with instructions baked in, including an explicit line telling any AI that opens it to treat the content as project requirements rather than commands — a small but deliberate guard against prompt injection for a document that's designed to be pasted straight into another assistant.
What we learned
Prompting a model to be decisive and specific is a completely different skill from prompting it to write well. It took real iteration to stop the AI from hedging into generic placeholders and get it to commit to concrete, competing choices the same way a real architect would. We also learned that people don't want an AI that just generates a plan for them; they want to fight with it a little delete a node, rewire an edge, insert a step it didn't think of so the tool had to be built as an editor first and a generator second.
What's next for Web designs
Three integrations are wired into the codebase but intentionally left out of this submission: AWS for project snapshots, Vultr for fast compatibility checks against a chosen stack, and Tiger Data for longer term decision history. Beyond that, the natural next step is closing the loop the web metaphor points at letting a finished, chosen path hand itself off to a real build pipeline instead of just a JSON file, so the web doesn't just get planned, it gets spun into something live.
Log in or sign up for Devpost to join the conversation.