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

Every year, freelancers suffer significant financial losses due to friendly fraud. Banks often reject WhatsApp or Slack screenshots as proof of work, making it challenging for freelancers to defend themselves. To address this issue, I developed Nyaya AI. This innovative solution cross-references initial contracts with informal chat logs, automatically generating a Visa/Mastercard-compliant legal rebuttal.

Given the potential dangers of AI hallucinations in the FinTech sector, I had to build a highly constrained, privacy-first architecture. The system is designed to be zero SDKs and serverless, built on Next.js using native Node.js fetch to communicate with the LLM (qwen-3.8-27b). This keeps the edge footprint tiny and cold starts fast.

For added security, the system is secured with Upstash Redis IP rate-limiting. To ensure data privacy, In-Memory PII Redaction is implemented, where zero databases are used. Before raw chat logs reach the LLM, an ephemeral regex buffer strips out sensitive information such as credit cards, SSNs, and emails. To mitigate the risk of AI hallucinations, an Anti-Hallucination Guardrails is in place.

The prompt uses a Verbatim Citation Mandate, requiring the AI to anchor every claim to a specific timestamp in the logs. If proof is missing, the AI refuses to guess and explicitly throws a Missing Documentation Warning. Given the financial nature of this application, the focus was on building a strict, verifiable pipeline rather than allowing for AI creativity.

I welcome feedback from the community regarding the architecture and potential edge cases. You can find the GitHub repository at https://github.com/ursarthak001/NYAYA-AI and the demo at https://nyaya01.vercel.app/.