Prompt Compass - Cut your AI bill up to 96% with a 0.57 MB prompt router

Hi Product Hunt Here's the problem I got obsessed with: my own AI bill. Most of what every LLM app pays frontier-model prices for is "what's the weather" in a suit prompts a 10–30× cheaper model handles identically. Routing fixes that, but the standard routing stack (~88 MB embedding model + vector DB) costs enough in size and latency to eat its own savings. So I spent months shrinking the route decision into **one 0.57 MB model** , smaller than a single photo that answers in **~5 ms on a plain CPU**. On assistant-style traffic it cuts the bill **up to 96%**. And since it already reads every prompt, the same call throws in what others sell separately: PII blocking and jailbreak detection. The firewall is the free bonus. What makes it different: 🪶 154× lighter than a typical embedding-based router ⚡ Decides on-device — your prompts can skip the network entirely 🌍 EN · ES · FR · DE · HI, and disguised/obfuscated text doesn't sneak past 🔒 The hosted API returns a decision and discards the prompt — never stored, never trained on 📊 Real numbers, published with methodology: ~82% on 2,022 held-out prompts (leakage-checked), 87.5% of real-world jailbreaks caught, <5% false blocks. I'd rather show an honest 82 than a suspicious 99. Free tier with no card, an npm SDK that runs fully on-device, and a VS Code / Cursor extension.