Engineeringemerging

AI Evals Engineer

Also hired as: Evaluation Engineer · Model Quality Engineer · AI Reliability Engineer

Every serious AI product now employs someone to answer one question: is it actually good?

US salary · 2026
$150k – $300k
A new title, so bands borrow from senior SWE and ML engineering; lab eval teams and staff-level roles exceed this. Treat as directional 2026 estimates.
Typical background
SDET and QA engineers leveling up, backend engineers who owned quality, and data analysts who like building pipelines more than decks.
This roadmap
5 stages · 24 nodes
5 proof-of-work checkpoints
What is a AI Evals Engineer?

An AI evals engineer builds the measurement layer for AI products: golden datasets, LLM-as-judge pipelines, regression gates in CI, and the observability that tells a team whether the product got better or worse this week. Traditional testing assumes deterministic outputs; AI products produce different valid answers on every run, so the discipline had to be reinvented from scratch. You are the person who replaces 'it feels better' with numbers a team can ship against — and the person who catches the regression before customers do.

The role emerged because companies got burned. Teams shipped AI features on vibes, watched a model upgrade silently break their prompts, and had no way of knowing until complaints arrived. By 2025 the fix had a name — evals — and a maturing toolchain in Braintrust, LangSmith, and Promptfoo, and labs plus AI-forward companies started hiring engineers whose entire job is quality measurement. The 'SDET of the AI era' framing is directionally right, with one correction: evals gate launches, model choices, and upgrades, which puts this role much closer to the decisions that matter.

What distinguishes it from its neighbors: agent engineers build the system, you prove it works — and the proof is a product of its own, with datasets to curate, judges to validate against human labels, and dashboards leadership actually reads. The craft mixes engineering (CI, tracing, pipelines) with applied statistics (variance, sample sizes, judge bias) and taste (knowing which failures matter). QA and SDET engineers who make this jump often find it is the highest-leverage career move available to them.

What you'll actually do
  • Design golden datasets from production traffic, support tickets, and deliberately hard edge cases.
  • Build LLM-as-judge pipelines and validate the judges against human labels with agreement stats.
  • Wire regression gates into CI so no prompt or model change ships without passing evals.
  • Run model-upgrade bakeoffs and write the recommendation memo leadership acts on.
  • Build observability: trace capture, failure taxonomies, and drift dashboards.
  • Sit with domain experts to turn 'good output' into rubrics a judge can score.
  • Maintain adversarial test sets — injection, PII leaks, jailbreaks — alongside quality sets.
  • Report quality metrics honestly, including the weeks when the news is bad.
This role fits you if
  • You are the person who asks 'how do we know it works?' and means it.
  • Building infrastructure other engineers rely on satisfies you more than shipping features.
  • Variance, sample size, and rater bias are concepts you enjoy rather than endure.
  • Reading 100 failure transcripts sounds like detective work, not drudgery.
  • You have QA, SDET, or data experience and want the AI-era version of your job.
The toolbox
PythonSQLBraintrustLangSmithPromptfooLLM-as-judge designGolden datasetsGitHub Actions / CITracing & observabilityError analysisStatistics (variance, agreement)Annotation ops
The roadmap — 5 stages, 24 nodes
Stage 1

Why AI Breaks Testing

Weeks 1–3

Understand the problem before the tooling: nondeterministic outputs, and the error-analysis habit everything else builds on.

Why assertEquals dies hereCore
Same input, different valid outputs — and 'valid' needs a rubric. Learn why traditional test suites go green while the product gets worse.
Just enough model internalsCore
Sampling, temperature, and context limits — the mechanics behind output variance. You need this to know what is noise and what is regression.
Error analysis as a practiceCore
Read outputs one by one and name what went wrong: hallucination, format break, tone miss, refusal. Open coding like this is the single highest-leverage habit in the field.
Checkpoint: label 100 outputs✓ Checkpoint
Pull 100 real outputs from any AI app — yours or open — and hand-label them into a failure taxonomy. Publish the taxonomy, the counts, and the three failure modes you would fix first.
Stage 2

Eval Construction

Weeks 4–8

Golden sets, metrics, and judges — the core artifacts of the trade, built properly once so you can build them fast forever.

Golden datasetsCore
Source from real traffic, cover edge cases deliberately, and version like code. Learn the contamination trap: never let eval cases leak into prompts or training data.
Choosing metricsCore
Exact match, semantic similarity, rubric scores — each is right for a different task shape. Extraction wants exact match; summarization needs a rubric; know why.
LLM-as-judge, done properlyCore
Write rubrics a judge can apply, then measure the judge itself: position bias, length bias, self-preference. A judge you haven't validated against human labels is just vibes with extra steps.
Rubrics with domain expertsElective
Sit with a lawyer, doctor, or support lead and translate their quality bar into scorable criteria. This skill is what separates eval engineers from script writers.
Checkpoint: open-source an eval harness✓ Checkpoint
One task, 200+ cases, an LLM judge validated against your own human labels with an agreement stat, and a README explaining where the judge disagrees with you.
Stage 3

Evals in the Pipeline

Months 2–4

An eval nobody runs automatically is a doc. Wire measurement into CI and production so quality becomes enforceable.

Regression gates in CICore
Eval runs on every prompt or model change; merges block on score drops. Handle variance with multiple runs and thresholds so the gate is trusted, not overridden.
Production eval loopsCore
Sample live traffic into eval sets, score continuously, and alert on drift. The golden set tells you about yesterday's failures; production loops catch tomorrow's.
Tracing and observabilityCore
Instrument an app so every generation carries its prompt version, model, latency, and cost. Failure analysis without traces is archaeology.
Survey the tooling landscapeElective
Run the same eval in Braintrust, LangSmith, and Promptfoo, then form an opinion on build vs buy. Interviewers ask; have an answer with reasons.
Checkpoint: a CI gate that catches a regression✓ Checkpoint
Wire your harness into CI for a demo app so a bad prompt change fails the pull request. Then deliberately regress the prompt and screenshot the red build — proof the gate works.
Stage 4

Agent Evals and Adversarial Testing

Months 4–6

Agents and attackers break simple evals. Grading trajectories and testing for injection is the frontier of the discipline.

Trajectory evalsCore
Grade the path, not just the answer: right tools, sane order, acceptable cost, graceful failure. Multi-step tasks need step-level and end-to-end scores.
Adversarial test setsCore
Injection via tool results, PII extraction attempts, jailbreak variants — encoded as regression cases so yesterday's exploit can never ship twice.
Model-upgrade bakeoffsCore
Run candidate models through your full suite, include cost and latency, and write the recommendation memo. This is the eval engineer's most visible deliverable inside companies.
Cost and latency as qualityElective
A 2% quality win at 3x cost is usually a loss. Add both as first-class eval dimensions and report them together.
Checkpoint: publish an agent-eval suite✓ Checkpoint
30+ multi-step tasks with trajectory grading, injection and PII cases included, run against two models — plus a comparison memo recommending one, with numbers.
Stage 5

Own Quality Somewhere

Months 6–8

Turn the skill into a mandate: prove ROI on a real system, then interview with case studies instead of claims.

The eval engineer pitchCore
You are the reliability layer for AI products: every caught regression is a shipped incident that didn't happen. Frame your work in prevented-incident terms — that is the language budgets speak.
Interview preparationCore
Expect 'design an eval system for X' plus live error analysis on sample transcripts. Your taxonomies and harnesses are rehearsal material; practice narrating them.
Community proofCore
Contribute cases to open eval suites, publish a dataset, or write the definitive post on judge validation. The field is young; visible work compounds fast.
Pick your doorElective
Lab eval teams, eval-tooling vendors, or enterprises standing up their first AI quality function. The third group is the biggest and least competitive entry point.
Checkpoint: a real engagement✓ Checkpoint
Stand up an eval suite for an OSS project or a team's AI feature and get it merged or adopted. The pull request plus a caught-regression story is your interview case study.
Build your portfolio

Nobody hires a AI Evals Engineer off a certificate. They hire off proof. Ship these and put them where people can click them:

01
A failure taxonomy from 100 labeled outputs
Hand-label 100 real outputs from any AI app into a named failure taxonomy with counts — and the three failure modes you would fix first.
Proves: You can do error analysis — the field's highest-leverage habit and the first thing practitioners check for.
02
An eval harness with a judge-agreement study
One task, 200+ cases, an LLM judge validated against your own human labels with an agreement stat, and a README documenting exactly where the judge disagrees with you.
Proves: You treat a judge as an instrument to calibrate, not an oracle to trust.
03
A CI gate that caught a regression
Your harness wired into CI on a demo app, then a deliberately regressed prompt and the red build screenshot to prove the gate works.
Proves: You can make quality enforceable, not just observable.
04
An agent-eval suite with a bakeoff memo
30+ multi-step tasks with trajectory grading, injection and PII cases included, run against two models — plus a comparison memo recommending one, with cost and latency in the numbers.
Proves: You can evaluate the hard cases and turn measurement into a decision leadership can act on.
05
An adopted suite with a caught-regression story
An eval suite stood up for an OSS project or a team's real AI feature — merged, running, and credited with at least one regression caught before users saw it.
Proves: Your work survived adoption by people who didn't have to say yes — the case study that closes interviews.
Position your profile
Headline formula

I build the measurement layer for AI products: eval harnesses with CI gates and human-validated judges — [N] regressions caught before users saw them.

Resume bullets to earn
  • Built the eval harness for [product or OSS project]: [N]+ cases, judge at [X]% agreement with human labels, merged and running in CI.
  • Blocked [N] quality regressions pre-merge with CI eval gates, including the [model upgrade] that silently broke [feature].
  • Ran the [model A] vs [model B] bakeoff across [N] tasks with cost and latency included; the memo drove the team's migration decision.
  • Raised judge-human agreement from [X]% to [Y]% by correcting position and length bias in the rubric.
  • Maintain the adversarial regression set — [N] injection and PII cases — so no known exploit ships twice.
Where to be visible
  • Open-source your harnesses and datasets; the field is young enough that one well-documented eval suite makes you findable.
  • Publish the judge-validation study: agreement stats, bias corrections, where the judge still fails. Almost nobody writes this post, and practitioners bookmark it.
  • Contribute cases to open eval suites and file eval-shaped issues on OSS AI projects — small PRs, high signal.
  • Tell caught-regression stories, sanitized: the upgrade that silently broke extraction and the gate that caught it. Prevented incidents are the genre hiring managers read.
  • Live in the Braintrust, LangSmith, and Promptfoo communities — the users of those tools are exactly the teams hiring this role.
What interviews actually test
  • 'Design an eval for X' where X is fuzzy — helpfulness, brand voice, summary quality. Practice decomposing a vague bar into scorable criteria out loud.
  • Live error analysis on sample transcripts: name the failures, group them, prioritize. Your labeling reps are the direct preparation.
  • Judge-validation depth: how do you know the judge is right? Agreement stats, bias types, and spot-check protocol should come out in full sentences.
  • Statistics under questioning: run-to-run variance, sample size, when a two-point delta is noise. Instincts with numbers attached, not proofs.
  • The case-study walkthrough: a suite you built, the regression it caught, the decision it drove. One true story with numbers beats every hypothetical.
Who's hiring
AnthropicOpenAIBraintrustLangChain (LangSmith)Scale AISierraHarveyGalileoEnterprises shipping AI (fintech, legal, health)
Free courses that map to this role
Related roles
FAQ

Is AI evals engineer a real job title?

Increasingly, yes — labs, eval-tooling vendors, and AI-forward enterprises post it directly. It also hides inside 'AI engineer' and 'ML engineer, quality' postings where evals are half the job description. The title is young; the function is now standard.

Is this just QA or SDET rebranded?

There is honest overlap — test design, CI, quality ownership — but the leverage is different. Evals gate model choices, launches, and upgrades, involve statistics and dataset curation, and sit near decisions executives care about. Compensation and status reflect that difference.

Can LLM-as-judge actually be trusted?

Yes, for many tasks — if you validate the judge against human labels, measure agreement, and control for position, length, and self-preference bias. An unvalidated judge is just automated vibes. The validation work is precisely why this is an engineering role.

Do I need machine-learning knowledge?

Model internals, no — you evaluate through APIs. Statistics, yes: variance, sample sizes, and inter-rater agreement are daily tools. Comfort with Python, SQL, and CI matters more than any ML theory.

How is this different from red teaming?

Red teaming is adversarial discovery — finding new ways systems fail. Evals are continuous measurement — making sure known failures stay fixed and quality holds. In practice you convert red-team findings into regression cases, so the two functions feed each other.

Start building, today

Every stage above maps to free lessons on this site. No signup, no paywall — open the first course and ship your first checkpoint this week.

Browse the courses →