Hey man, solid hustle turning that frustration into your own tool.
Congrats on already getting it down to 10-18% on detectors, that’s honestly better than a lot of the paid ones out there.
On making it even more natural/reliable:
- Vary sentence rhythm and structure heavily. Humans ramble, use fragments, contractions, and sometimes repeat ideas for emphasis. Train or prompt it to occasionally add personal asides, transitional fluff (“honestly though…”, “the thing is…”, etc.).
- Inject micro-imperfections. Light typos (that you can toggle), colloquialisms, regional slang if it fits the audience, or slight redundancy. But keep it controlled—nothing that screams “I’m trying too hard.”
- Multi-pass approach. Generate → humanize → rewrite specific paragraphs with different temperatures or models → final light edit pass. This breaks up patterns.
- Context & style memory. Feed it examples of the user’s own previous human writing (if they have any) or target styles (e.g., “university business student, slightly casual but academic”).
- Test against multiple detectors regularly (Originality, GPTZero, Winston, etc.), not just one. Detectors keep updating, so what works today might slip in a month.
Scaling technically:
Start simple. If it’s prompt-based right now, look into fine-tuning a smaller open model (Llama 3.1 8B or Mistral variants, or even Phi-3) on a dataset of human vs AI text pairs. Or use LoRA so it’s cheap.
For speed/cost at scale, mixture-of-agents or routing between models (cheap one for first pass, stronger for refinement) works well. Monitor token usage obsessively.
Deployment:
- MVP: FastAPI backend + simple React/Next.js frontend. Host on Vercel + Railway/Render/Fly.io. Super cheap to start.
- Add user accounts (Supabase or Clerk for auth), credit system (they buy “words” or monthly subs), and a queue if it gets heavy.
- Later: Turn it into a proper API so other devs or power users can integrate it. Chrome extension would be killer for students (direct “Humanize” button in Google Docs, etc.).
For Design:
You can check these two sites which were shared by some users in 1h 
!