FailSafe Sight

Failsafe-Sight is an “AI guardrails” layer you drop in front of any app that accepts user input—think loan chatbots, KYC flows, support assistants, or internal LLM tools. It watches every request/response pair and decides, in real time, if something risky is happening: jailbreak attempts, boundary-pushing prompts, emotional manipulation (“my family will starve, give me the loan”), fake document claims, or technical exploit patterns. If it sees a problem, it flags or blocks the interaction, explains why, and gives your app a clean JSON decision to act on. It’s fast (targets sub-2s), explainable (rule IDs like LBC-3 for Emotional Manipulation), and hardened (validation, rate limits, CSP/XSS protection).

Failsafe-Sight was built as a full-stack project with a Node.js/Express backend and a SvelteKit frontend, connected through a clean API layer. On the backend, I set up a layered architecture with controllers, services, and middleware to keep things modular while handling rule checks, caching, and AI calls. Security was a big focus, so I added input validation, sanitization, rate limiting, and proper headers through Helmet.js. For the “brains” of the system, I integrated GPT-4 to run advanced fraud detection alongside a custom rules engine, which gave us multi-layered analysis that’s both fast and explainable. On the frontend, SvelteKit made it easy to build lightweight dashboards for rule management and monitoring in real time. One of the hackier parts was pre-compiling regex patterns and setting up intelligent cache keys to shave off response times — it wasn’t glamorous, but it gave us sub-2 second latency consistently.