Ask me about Derek

How this Q&A resists abuse

A public, unauthenticated LLM endpoint is an attack surface. These guardrails are as much part of the work as the Q&A itself:

• Grounded in the resume

  Designed to answer only from retrieved resume chunks; when something is not in them, it is instructed to say so rather than make things up.

• Off-topic task refusal

  Code writing, translation and other chores are designed to be declined, even when mixed into an otherwise legitimate question.

• Prompt-injection defenses

  Visitor input is fenced and smuggled fence tags are neutralized; chat history is treated as untrusted and sanitized first.

• Privacy boundaries

  Salary, phone and address are off limits; only contact info the owner chose to publish is shared. Q&A logs store no IPs.

• Rate & cost limits

  Daily global and per-IP caps, input length limits, oversized payloads rejected outright, CORS allowlist.

• Graceful degradation

  Clear errors when AI or retrieval fails; query rewriting falls back on timeout; the streaming UI degrades gracefully.