Ayman Khan

Built production Redis agent with a three-agent pipeline (parser → security guard → executor/formatter) and RedisClientManager featuring LRU eviction, TTL expiry, and background cleanup, reducing manual Redis operations for the team and automating more than 80% of the routine Redis operations.
Architected an Alert Triage & Notification System orchestrating three remote sub-agents via the A2A protocol to auto-generate incident tickets and notify affected users, reducing incident response initiation time from ~30 minutes to under 2 minutes.
Migrated Super Agent codebase from unofficial python-a2a to Google's official A2A SDK, implementing Agent Skills for Rakuten's core business logic and ensuring long-term maintainability.
Designed and built Agent Builder, a meta-agent system on Google ADK that autonomously generates production-ready agents from natural language descriptions using a Planner → Architect → Codegen → Validator pipeline with swappable per-stage skill files.

Built a unified multi-domain hallucination benchmark spanning 400 adversarial prompts, evaluating GPT-4o, Claude Sonnet, Gemini 2.5, and DeepSeek-V3 across factual, reasoning, and domain-specific tasks.
Designed the Hallucination Degree (HD) metric to quantify both frequency and severity of fabricated or irrelevant LLM outputs, enabling standardised cross-model comparison.
Prepared research draft and presentation for Samsung PRISM mentors; work contributed towards a publication.