eye-eye-ai

Dropping into an unfamiliar codebase is one of the most expensive moments in software. The code is all there, but the *architecture* — the systems that actually exist, how a request becomes a response, what is verified and what is not — lives in nobody's head and nowhere on disk. The tools we reach for each show a slice and miss the shape: What's missing is a **spatial, concept-first model** of a codebase that a human can inspect, navigate, and narrate — something that represents *runtime systems, verification structure, and user-understandable flows* rather than folders. **What we built.** This project turns a real repository into an explorable **3D architecture mission map**. Gemini 3.5 Flash acts as a fast *architecture scout*: it reads the repository as raw evidence (files, symbols, tests, docs, configuration) and synthesizes that evidence into higher-level architectural meaning, rendered as a polished sci-fi control-room scene. Crucially, files are **evidence, not identity** — the scene primarily shows architectural concepts, subsystems, and flows, with files demoted to grounding metadata behind them. The experience is organized around three authored **lenses**, each answering one question about the same codebase: | Lens | Question it answers | Vertical meaning | |---|---|---| | **System Map** | "What major parts exist, and how do they nest?" | abstraction depth | | **App Assembly** | "How does a developer's app become request handling?" | runtime progression | | **Test Scaffold** | "How is this behavior verified and explained?" | production vs. verification | Every lens reuses one small visual vocabulary — floor layers, vertical zones, component hulls, code blocks, scaffold rails, transit edges, and label cards — so the *same* mental model rearranges to tell three different stories. Users orbit the scene, fly with elevation, recenter, focus any element to read its description, and follow authored traversal paths and walkthroughs as a guided narration.