I built a serverless AI engine to fight freelance chargebacks. Looking for architecture feedback!

Hey everyone, I'm here to showcase my project that I just built for the BharatCode Build League, and I am looking for some technical feedback.

The Problem: Freelancers lose billions to chargebacks because banks reject messy WhatsApp/Slack screenshots as legal proof of work.

The Solution: I built NYAYA-AI, a Sovereign AI defense engine that cross-references your contract against your raw chat logs to instantly generate a Visa/Mastercard-compliant rebuttal dossier.

Because AI hallucinations in FinTech are dangerous, I built this with strict architectural constraints:

• ⚡ Zero SDKs: Next.js using native Node.js fetch (for qwen-3.8-27b) + Upstash Redis for fast edge serverless performance.

• 🔒 In-Memory PII Redaction: An ephemeral regex buffer strips sensitive data (SSNs, cards) before it hits the LLM. Zero databases.

• 🛑 Strict Anti-Hallucination: A "Verbatim Citation Mandate" mathematically forces the AI to anchor every claim to a specific timestamp, or it explicitly refuses to guess.

The repo is completely open-source. I would love for the engineers and founders here to roast my architecture:

• Is regex PII redaction too naive here?

• What edge cases did I miss for serverless edge functions?

💻 GitHub Repo: https://github.com/ursarthak001/NYAYA-AI