Responsible AI - Prototype
AI agent reference implementation HelloBot is a deliberately simple "hello world" agent (it greets people and can send a greeting by email). The agent stays tiny on purpose, so the focus is the enterprise controls around it : # Pillar Implementation Industry standard used 1 Guardrails (input + output) app/guardrails.py : blocks prompt injection, denied topics and over-long input. Redacts PII (SSN, card, Aadhaar, PAN , phone). Blocks system-prompt leaks. Every decision is tagged with a policy id Pluggable PIIDetector interface (drop in Microsoft Presidio / cloud DLP) 2 Observability, explainability, auditability app/telemetry.py : distributed tracing with GenAI semantic conventions ( gen_ai.usage.* , gen_ai.request.model ). app/audit.py : every step goes into an append-only, SHA-256 hash-chained audit log, and tampering is detected. Raw input is stored only as a hash OpenTelemetry → OTLP → Jaeger (works with Datadog / Grafana / Azure Monitor) · SQLAlchemy 3 Grading / ...