AI Agent Evaluation Engineer
Posted Updated
Description
- Cresta’s AI Receptionist answers the phone for small businesses that would otherwise miss the call
- It handles millions of conversations a month today, making sure no call goes unanswered and helping owners complete the front-office tasks that keep their business running, and it does so across hundreds of customers
- Whether those conversations are actually good is the single most important question in the business, and it deserves a full-time owner
- That’s you
- As our first AI Agent Evaluation Engineer, you define what quality means for every agent we run, you build the systems that measure it across every conversation, and you are the person the company trusts when the question is “how good is it, really?”
- The work combines hands-on conversation analysis, rigorous evaluation design, and engineering execution: reading and annotating real conversations to learn how the agent actually fails, building the automated evaluation systems (LLM-based judges, golden sets, regression suites) that measure those failures at scale, and keeping those systems honest by calibrating them against human judgment
- Fixing what you find is welcome but not the core of the job; an accurate, prioritized picture of agent quality is
- Define and maintain the evaluation rubric for the agent across three axes: adherence (did it follow its design, call the right tools, respect guardrails), outcome (did the conversation achieve its goal), and conversation experience (latency, turn-taking, tone, repetition)
- Write the annotation guidelines and failure taxonomy the team uses to talk about failures precisely; run calibration sessions and measure inter-rater agreement
- Classify and prioritize failure modes by severity and frequency so engineering and product work on what matters most
- Review production conversations by hand every week, leverage QA partners; open-code failures, cluster them, and turn the clusters into the evaluations worth automating
- Trace each failure to its likely cause (prompt, knowledge gap, tool or flow bug, speech recognition) and route it to the right owner with enough evidence to act on
- Evaluate across languages and report quality by segment (language, accent, audio quality, intent), not just in aggregate
- Build and operate the evaluation pipeline that scores every production conversation, combining automated checks, LLM-based judges, and human review
- Calibrate every automated judge against human labels before trusting it; treat judge error as the primary measurement risk
- Design sampling strategies (stratified by intent, language, caller type, audio conditions) so that reported changes in quality are real
- Build and maintain the golden set and regression suite that every prompt, model, knowledge, and flow change is validated against before release
- Maintain the failure-mode coverage map: which behaviors are tested, partially tested, or unmeasured; close the gaps
- Design evaluations for behaviors no standard harness covers: high-stakes conversation types such as handoffs to a human, misuse and unwanted-caller handling, and resilience under upstream failures
- Define the quality and latency thresholds the agent is held to
- Report quality trends to internal stakeholders and customers clearly, with uncertainty stated and without burying bad news
- Partner with forward-deployed engineers, backend engineers, product managers, and the voice/language engineering team to turn findings into prioritized recommendations
- Document methodology so results are reproducible and the evaluation practice outlives any one person
Benefits
- Weekly lunch & learn
- Work with an A+ team of ex-Founders, world leading scientists, and nationally competitive athletes
- Mentorship - A world class mentorship board advises and mentors our team-members to capture the opportunity ahead of us
- Strong programming skills in Python and experience with LLM APIs; able to build and ship evaluation tooling independently. Fine-tuning and ML frameworks are not required
- Strong written communication; you author the guidelines others follow
- Deep understanding of evaluation methodology: what golden sets, judges, regression suites, and calibration are for, and when each fails. Tool-agnostic; you will work in our in-house tooling
- Built and operated an end-to-end evaluation system for a production LLM product (agent or otherwise): rubric, human labels, automated judge, regression suite
- Fluent in how LLM agents work: prompts, tool calls, context handling, multi-turn state. You can read a trace and tell what the agent was doing and why
- Able to write a rubric that independent reviewers would score the same way, and to measure whether they do
- 3+ years of professional experience in software engineering, applied ML, or data-intensive research roles, including at least 1–2 years working hands-on with LLM-based systems in production
- Able to design a test from an open-ended quality question: you decide what to measure and how, rather than executing a predefined plan
- Professional fluency in at least one language other than English, sufficient to evaluate multilingual agent behavior directly
- Willing to do non-scalable work weekly, both yourself and by directing our QA partners: listening to calls, labeling conversations, building a failure taxonomy from scratch
- Bachelor’s or Master’s degree in Computer Science, Computational Linguistics, Cognitive Science, Statistics, or a related field, or equivalent practical experience
- We highly value your time and engineering craftsmanship. To ensure a streamlined and objective process, the first step for qualified applicants following an initial resume review is a self-paced CoderPad technical assessment. This allows you to showcase your practical coding and problem-solving skills in a real-world environment before moving into conversational interview rounds
Skills
Sign in to read and join the discussion.