Built a Sovereign AI to Fight FinTech Chargebacks (Zero SDKs, Zero Hallucinations)

Freelancers lose millions every year to "friendly fraud" because banks reject messy WhatsApp or Slack screenshots as proof of work.

I built Nyaya AI to solve this. It cross-references your initial contract against your informal chat logs to automatically generate a Visa/Mastercard-compliant legal rebuttal.

Because AI hallucinations in FinTech are dangerous, I had to build a highly constrained, privacy-first architecture:

Zero SDKs & Serverless: Built on Next.js using native Node.js fetch to talk to the LLM (qwen-3.8-27b). This keeps the edge footprint tiny and cold starts fast. Secured with Upstash Redis IP rate-limiting.

In-Memory PII Redaction: Zero databases. Before raw chat logs hit the LLM, an ephemeral regex buffer strips out credit cards, SSNs, and emails.

Anti-Hallucination Guardrails: The prompt uses a "Verbatim Citation Mandate." The AI must anchor every claim to a specific timestamp in the logs. If proof is missing, it refuses to guess and explicitly throws a "Missing Documentation Warning."

Building for the financial sector means optimizing for strict, verifiable pipelines rather than AI "creativity."

I'd love some feedback from the community on the architecture and edge cases!

github: https://github.com/ursarthak001/NYAYA-AI demo: https://nyaya01.vercel.app/