I built a 0.57 MB model that cuts AI bills by up to 96% — solo, from India, $0 spent
Hey IH 👋
I'm Rishabh, a solo bootstrapped founder from India, and I just took my first product live: Prompt Compass — a 0.57 MB AI model that does four jobs in one ~5 ms call:
routes prompts (simple → cheap model, hard → expensive model)
blocks PII before it reaches any cloud
catches jailbreak / prompt-injection attempts
works in EN, ES, FR, DE, HI
Why this exists: every LLM app needs those four things, and the standard answer is four separate tools — an ~88 MB embedding router, a vector DB, a PII scanner, an injection classifier. I got stubborn about a question: can all four fit in something smaller than a photo? Answer: yes, with trade-offs I'll tell you about honestly.
The honest numbers (published, leakage-checked):
82% overall on 2,022 real held-out prompts
87.5% of real-world jailbreaks caught
<5% of genuine prompts wrongly blocked
~5 ms per decision on plain CPU, 0.57 MB total
Lakera and Azure likely beat my raw accuracy. They're also cloud-only, heavier by 100–1000×, and Lakera-class pricing starts around $500/month at scale. My niche: smallest, fastest, cheapest, most private — and the only on-device option. I publish the 82 instead of demo-rigging a 99, and that honesty is half the brand.
What "$0 budget" actually means:
marketing: this post, and others like it — all organic
team: me
Business model:
Free tier (evaluation + VS Code extension traffic)
Lite $9/mo → Pro $20/mo (individual devs)
Commercial from $79/mo + usage (production apps)
On-device SDK: Indie $49 / Business $199
Enterprise: custom
Revenue so far: $0 MRR. Monthly updates incoming, real numbers, including the embarrassing ones.
What I'd love from IH:
Brutal feedback on the pricing ladder
If you're building with LLMs — would you actually put this in the loop?
What would make you swap your current guardrails setup?
Try it free (no card): compass.mpcfintech.com
Made in India 🇮🇳
