Browser IDEs (Codespaces, StackBlitz, Cursor Web, Replit) trail native VS Code on two things: find-in-files latency and ranking the AI suggestions they already produce. The first is a search problem; the second is a compute problem.
The client-side search layer behind “find in files” (today single-threaded JS regex over an in-memory file corpus), and the ranker that sits between an LLM completion endpoint and the UI (today missing, or doing an extra cloud round-trip to a judge model).
Native-IDE-grade substring + multi-pattern search across a 20K-100K file repo, with regex / wildcard / fuzzy refused out loud (no silent fallback). And a local re-ranker that prioritises AI completions without spending another LLM call.
Hadron is not claiming the GPU is fastest. It picks the right backend for each operation. Substring search routes to GPU (GPU-accelerated search); multi-pattern routes to CPU multi-pattern path; regex / wildcard / fuzzy hit a categorical refusal (patent-protected). The candidate-ranker matmul + sort pick CPU or GPU per call based on input shape (patent-protected).