Generative AI Quality Engineer - Assistant Vice President

Design and execute end-to-end test strategies for Agentic AI pipelines, including single-agent and multi-agent workflows.

Validate agent reasoning, planning, and decision-making chains (e.g., ReAct, Chain-of-Thought, Plan-and-Execute, Reflexion).

Test tool-use correctness — ensuring agents invoke the right tools, with correct parameters, at the right time.

Evaluate agent memory systems (short-term, long-term, episodic) for accuracy and context retention across sessions.

Validate agent handoff and delegation logic in multi-agent orchestration frameworks (e.g., AutoGen, CrewAI, LangGraph).

Test termination conditions, loop detection, and infinite loop prevention in autonomous agent loops.

Design comprehensive test strategies for end-to-end RAG pipelines — covering ingestion, chunking, embedding, retrieval, reranking, and generation stages.

Validate retrieval accuracy and relevance — ensuring the correct context chunks are retrieved for a given query.

Test embedding model quality and vector similarity thresholds across different document corpora.

Evaluate faithfulness, groundedness, and answer relevance of generated responses using frameworks like RAGAS, TruLens, DeepEval.

Test chunking strategies (fixed, semantic, hierarchical) for their impact on retrieval quality.

Validate context window management — ensuring retrieved context does not exceed token limits or degrade generation quality.

Conduct end-to-end regression testing when the underlying knowledge base, embedding model, or LLM changes.

Test multi-turn conversational RAG for context coherence and citation accuracy across turns.

Build and maintain automated test harnesses for Agentic and RAG systems, including agent trajectory replay, tool mock injection, and prompt simulation.

Develop automated evaluation pipelines integrated into CI/CD workflows for continuous model and agent validation.

Create data validation and data quality frameworks (using Great Expectations, Deequ, or custom tooling) for training, retrieval, and inference data.

Build prompt regression suites to detect behavioral drift across LLM versions or prompt changes.

Implement determinism and reproducibility tests for stochastic LLM-based decisions.

Automate vector database validation — index integrity, embedding drift, and retrieval consistency checks.

Conduct red-teaming and adversarial testing to uncover jailbreaks, prompt injection vulnerabilities, and goal misalignment in LLM-based systems.

Test output guardrails and content filters for unsafe, biased, toxic, or out-of-scope model behavior.

Validate privilege escalation controls — ensuring agents do not exceed permitted actions or access unauthorized resources.

Perform data poisoning and backdoor attack simulations to assess model robustness.

Evaluate models for bias, fairness, and discrimination using frameworks such as AI Fairness 360 and Aequitas.

Test PII leakage and data privacy controls in RAG and agent pipelines in accordance with GDPR, CCPA, and internal data governance policies.

Conduct security testing aligned with the OWASP Top 10 for LLM Applications, including:

Prompt Injection (Direct & Indirect)

Insecure Output Handling

Training Data Poisoning

Insecure Plugin / Tool Design

Sensitive Information Disclosure

Validate constitutional AI constraints, RLHF-aligned behavior boundaries, and system prompt integrity.

Collaborate with cybersecurity teams on AI-specific threat modeling and vulnerability management.

Maintain safety testing playbooks and document red-team findings with severity ratings and remediation recommendations.

Define and execute load, stress, soak, and spike testing for AI-powered APIs, inference endpoints, and agent orchestration services.

Measure and optimize end-to-end latency across RAG and agentic pipelines — from query to final response.

Benchmark LLM inference throughput (tokens/second) and identify bottlenecks across model serving infrastructure.

Test auto-scaling behavior of AI services under variable load conditions.

Validate circuit breaker, retry, and fallback mechanisms in agentic and RAG systems for graceful degradation.

Test vector database performance — query latency, index build time, and retrieval accuracy under high concurrency.

Conduct cost efficiency analysis — measuring token consumption, API call costs, and infrastructure spend per agent task.

Establish SLOs (Service Level Objectives) and SLAs for AI system availability, latency percentiles (P50, P95, P99), and error rates.

Collaborate with MLOps teams to set up observability dashboards, monitoring alerts, and automated anomaly detection for production AI systems.

Perform chaos engineering experiments to validate agent and RAG system resilience under infrastructure failures.

Deep understanding of RAG architecture patterns — naive RAG, advanced RAG, modular RAG.

Solid grasp of agent design patterns: ReAct, Plan-and-Execute, Reflexion, MRKL, Mixture-of-Agents.

Familiarity with AI safety and alignment principles (RLHF, Constitutional AI, guardrail layers).

Knowledge of token economics, context management, and LLM cost optimization.

Proficiency in performance engineering methodologies for distributed AI systems.

Citi is an equal opportunity employer, and qualified candidates will receive consideration without regard to their race, color, religion, sex, sexual orientation, gender identity, national origin, disability, status as a protected veteran, or any other characteristic protected by law.

If you are a person with a disability and need a reasonable accommodation to use our search tools and/or apply for a career opportunity review Accessibility at Citi. View Citi’s EEO Policy Statement and the Know Your Rights poster.