PortPilot — CUDA → ROCm Readiness Score

It's software lock-in, not hardware, that keeps teams off AMD. Teams assume their CUDA-first codebase is "stuck on NVIDIA" — but most have never measured how true that is. The first question before any migration isn't how — it's how far am I from ROCm? Today that means a senior engineer auditing the repo by hand, or trusting an LLM's guess. PortPilot answers it in minutes, deterministically. How it works: Paste any public GitHub repo. A FastAPI backend scans it with Python ast analysis and regex rules, classifying every CUDA dependency into one of three buckets: A — Works as-is (torch.cuda.*, device="cuda" — runs on ROCm PyTorch unchanged), B — Mechanical change (cupy, flash_attn, triton, nvidia-* wheels — HIPify-style swap), C — Manual blocker (tensorrt, inline PTX, cuda-python driver API). The readiness score is a pure function of the tally: score = 100·(1.0A + 0.5B + 0.0C)/(A+B+C). Same repo in, same score out. Where the LLM fits — and where it doesn't: Fireworks AI on AMD Instinct is strictly an advisory layer — it explains each finding and drafts suggested diffs, labeled advisory and never auto-applied. The LLM never classifies and never touches the score. Leaders can trust the number; engineers can use the explanations. Why not another auto-migration agent? Auto-porting tools ask you to trust an LLM to rewrite code before you've decided to migrate. PortPilot sits one step earlier: the assessment product — a shareable readiness report for decision-makers. It also busts the lock-in myth: scan karpathy/nanoGPT and get 100/100 — already ROCm-ready. Shipped: live at portpilot-one.vercel.app (no sign-up), shareable reports, offline demo, benchmark page. Next.js/Vercel frontend; FastAPI backend on Railway. Containerized. Built solo. Market: every ML team weighing AMD Instinct needs this number first — the free top-of-funnel scan for the whole ROCm migration ecosystem.