Sai Haresh Anand S

Profile README

On-call incident agent with Hindsight memory and a provenance gate: a remembered fix must be re-verified before it runs

AI safety navigation app — LLM-scored route safety with Gemini → Groq fallback, SOS alerts, live trip sharing — plus an open adversarial reliability audit of the system itself

Autonomous multi-strategy trading engine with an independent risk control plane, drawdown kill switch, and staged backtest → paper → live promotion

Activation-space governance: detecting and suppressing a trained deceptive behavior at inference time, without modifying weights. Week-1 alignment research prototype.

Reclaim: failed-payment recovery agent for the Razorpay AI Buildathon. Rules for known failure reasons, an LLM only for the ambiguous ones.