Senior AI Evals Engineer

About the Role

The Senior AI Evals Test Engineer  is a hands-on technical role responsible for defining, developing, and executing robust evaluation frameworks and test strategies for AI agents built using the Google Agent Development Kit (ADK) or similar sort of agentic frameworks. This role bridges AI development and reliable deployment — ensuring our agents are safe, ethical, effective, and meet high-quality performance standards. A significant focus of the role is on Responsible AI and Safety Evals, including adversarial testing, red-teaming, bias detection, and toxicity measurement. The engineer will design scalable evaluation pipelines, build synthetic test environments, and work closely with AI/ML Engineers, Researchers, and policy teams to ensure agents are production-ready and compliant with regulatory and ethical standards. The role is 70% Automation and 30% Manual Testing, and sits within the Applied AI: Test Engineering team as a key technical authority.

Key Responsibilities

Design and build agent evaluation frameworks — create end-to-end infrastructure for evaluating LLM-based agents across correctness, latency, cost, safety, and task completion rate

Build synthetic testing environments and simulation strategies to stress-test AI agents under diverse real-world conditions, with specific focus on agents built on Google ADK and related Google AI/ML services (e.g., Vertex AI, Gemini models)

Build golden datasets and test harnesses — curate domain-specific evaluation sets for Rubrik's product verticals (data security, SaaS protection, cyber recovery, compliance)

Establish eval-as-a-service — build and enhance centralized evaluation platform that any IT team members can self-serve to test their agents against standardized and custom benchmarks

Automate regression and drift detection — implement CI/CD-integrated eval pipelines that catch regressions on every model or prompt change before production deployment

Define eval taxonomies and rubrics — establish structured scoring criteria for agentic workflows (tool use accuracy, multi-step reasoning, hallucination rate, policy adherence, error recovery)

Instrument production observability — build real-time monitoring for agent quality in production: success rates, user corrections, fallback triggers, escalation patterns

Benchmark and select models — run structured model comparisons (cost, quality, latency trade-offs) to inform model selection decisions across agent use cases

Analyze and interpret evaluation results, providing clear, actionable quality and safety reports to stakeholders and development teams, with specific focus on risk metrics, trend analysis, and mitigation recommendations

Define comprehensive end-to-end QA strategies — functional, integration, regression, and user acceptance testing (UAT) — tailored for conversational and goal-oriented AI agents with tool-use, memory, and multi-turn reasoning capabilities

Develop and maintain detailed test artifacts (test plans, test cases, test scenarios, automated test suites) covering agent features, tool invocation, memory persistence, and reasoning accuracy across evolving model versions

Identify, document, prioritize, and track defects, performance regressions, and alignment failures using Jira, collaborating closely with Applied AI Engineers to analyze root causes and validate fixes

Evaluate new technologies and patterns (e.g., new tools for Agent Evals), run proofs-of-concept, and influence the platform roadmap

Partner with policy and legal teams to ensure agent evaluations systematically cover regulatory compliance, fairness objectives, and emerging AI governance requirements

Provide technical leadership to engineers, including solutioning, design reviews, code reviews, pair programming, and mentoring of junior and contract engineers

Collaborate cross-functionally — work closely with ML engineers, product teams, platform engineers, and security to embed eval-driven quality gates into the agent development lifecycle

Required Qualifications

Education & Experience

Bachelor's or Master's degree in Computer Science, Information Systems, Engineering, or equivalent practical experience

8+ years of overall software Test engineering experience, with 5+ years focused on test engineering, QA automation, or developer tooling, and demonstrated ability to build evaluation systems across multiple AI/ML domains, conversational agents, or Large Language Models (LLMs)

2+ years working with ML/AI systems, building, testing, or evaluating models in production environments, including hands-on experience with LLM-based applications, prompt engineering, RAG pipelines, or agentic workflows

2+ years as a senior/lead, owning evaluation strategy, quality frameworks, and technical decisions for complex AI agent systems in an enterprise product environment

1+ years in designing or operating evaluation frameworks specifically for generative AI / LLM agents, including benchmarking, red-teaming, and production quality monitoring in the AI ADLC model

Technical Skills

Deep hands-on experience with LLM evaluation — familiarity with frameworks and libraries such as LangSmith, DeepEval,LLM as a Judge, Promptfoo, RAGAS, Hugging Face evaluation tooling, custom eval harnesses

Strong understanding of agentic AI patterns — tool use, function calling, multi-step planning, ReAct/CoT reasoning chains, retrieval-augmented generation (RAG)

Production experience with eval pipelines — building automated, CI-integrated evaluation infrastructure (not just notebooks)

Proficiency in Python — strong scripting and tooling skills; experience with pytest, async patterns, and data processing at scale

Familiarity with Google Cloud Platform (GCP) or any Cloud Service relevant to AI/ML (e.g., Vertex AI) and experience integrating testing into MLOps workflows

Statistical literacy — ability to design experiments, interpret eval metrics with appropriate confidence, and avoid common pitfalls (overfitting to benchmarks, metric gaming)

Enterprise product sensibility — understanding of how AI quality translates to customer trust in a B2B/enterprise context (security, compliance, reliability SLAs)

Excellent communication — ability to present eval results and quality trade-offs to engineering leadership, articulate risks and safety findings to non-technical audiences, and drive alignment on ship/no-ship decisions

Hands-on experience with CI/CD pipelines and Git-based workflows, including automated testing, static analysis, deployment automation, and environment configuration for our DevTestOps

Tech Stack:

Languages: Python(Must), TypeScript, Java (Optional)

Test Automation:Selenium, Playwright

LLM Providers: Anthropic Claude, OpenAI, Azure OpenAI, open-source models

Eval Tooling: DeepEvals(Must), LLM as a Judge (Must), Promptfoo/RAGAS/LangSmith/ Custom harnesses

Agent Evals Frameworks/libraries: LangGraph, Claude Agent SDK, custom orchestrators

Infrastructure:DevOps (Basic), Kubernetes, CI/CD (Jenkins/GitHub Actions), observability stacks (Datadog/Grafana)

Preferred Qualifications

Deep Learning/NLP Knowledge: Fundamental understanding of how LLMs work, including concepts like RAG, prompting, and the technical challenges related to alignment

Data Analysis: Experience with data analysis and visualization tools (e.g., Pandas, SQL) for analyzing evaluation and safety results

Policy Background: Familiarity with AI ethics principles and industry standards for responsible AI development

Garbage-in/Garbage-out (GIGO) Tools: Any tool or framework that specializes in testing the model's response to highly ambiguous, misleading, or malformed inputs to test for robustness and safety failures

Experience testing complex APIs, especially for tool-integration and function-calling scenarios in agentic workflows

Experience with prompt injection (Must) and data poisoning skills is a plus

Certifications in AI/ML, cloud platforms (GCP, AWS), AWS Machine Learning Specialty, Google Professional ML Engineer, DeepLearning.AI specializations and software testing methodologies

Soft Skills

Proven ability to lead without direct authority, influencing cross-functional teams and driving technical decisions through clarity and data rather than hierarchy.

Experience working in agile environments, collaborating closely with product owners, scrum teams, and distributed stakeholders across time zones.

Excellent communication and documentation skills, comfortable presenting solutions and trade-offs to engineers, managers, and senior business stakeholders.

Strong architectural thinking and problem-solving skills, including the ability to decompose complex business processes into simple solutions

Nice to have

Experience evaluating agents in cybersecurity, data protection, or infrastructure domains

Familiarity with Rubrik's product portfolio or similar enterprise data management platforms

Hands-on with human-in-the-loop eval workflows (annotation platforms, inter-rater reliability)

Experience with model fine-tuning evaluation (pre/post quality comparison)

Background in building internal developer platforms or tooling teams

Contributions to open-source eval frameworks or published work on AI evaluation methodology

Drive red-teaming and adversarial testing — design adversarial evaluation suites for jailbreak resistance, data leakage prevention, and boundary enforcement in enterprise-grade agents