ML Annotation QA Engineer

We are looking for an ML Annotation QA Engineer to own the quality of annotated data across our computer vision and machine learning programs. This role is responsible for the judgment-heavy analysis that cannot be reliably outsourced, for the decision rules behind it, and for turning annotation output into an ongoing read on how our systems are actually performing in the field.

We work with an external annotation partner at production volume, and that continues. What we need in house is someone who can analyse the annotated data, make and defend the calls the vendor cannot make consistently, and build the aggregate view that shows which facilities and equipment are degrading and why. Annotation drift, a model regression, a tool bug, and genuine field degradation all look similar in a chart and require completely different responses — telling them apart is the core of the job.

You will work closely with Machine Learning Engineers, QA, and Engineering, and the first assignment is our warehouse forklift vision program, where barcode readability and localisation analysis are the immediate need. From there the remit grows with us: drone imagery annotation today, and new task types as customer-driven capabilities come online. Success in this role requires a combination of analytical rigor, sound judgment under ambiguity, and clear written communication.

Own the judgment-heavy quality analysis on annotated data that cannot be reliably outsourced — working the daily review queue and producing verdicts and root cause in house.

Own, version, and refine the verdict taxonomy, decision rules, and quality guidelines for the categories you cover.

Build and maintain performance trackers over annotated data — error rates by facility, site, equipment, and data format over time, against an agreed baseline.

Detect anomalies against that baseline and flag them the day they appear rather than weeks later.

Run root cause analysis on flagged anomalies, distinguishing annotation error from model or system error from genuine degradation in the field.

Report findings to engineering and ML with reproducible evidence and stated confidence, fast enough that the issue is still observable.

Identify systematic failure patterns rather than one-off misses, and maintain a documented pattern library others can use.

Query and analyse annotation data directly with Python and SQL to test hypotheses, without waiting on extracts from anyone.

Feed annotation-quality findings back as concrete SOP and instruction changes when the root cause is labeling rather than system behaviour.

Specify annotation tool improvements — what the tool should surface so analysis stops requiring manual work — and validate the fixes.

Stand up quality analysis and reporting for new annotation programs as they come online.

Track work across Jira and contribute to pre-release validation for the behaviours you cover.

Within your first three months, you will be expected to:

Take over barcode and location root cause analysis from the annotation vendor.

Move from supervised review to owning the daily queue at the agreed review threshold, inside the expected time budget.

Become the primary owner of verdicts and root cause for the categories you cover.

Develop working expertise in the data and the domain behind it:

What is captured, at what grain, and where the known quality limits are

Racks, locations, levels, bins and bin types; LPN, SKU, AWB and other code formats; exceptions and exception types; OCR versus barcode and the failure modes of each

Stand up a facility performance tracker with an agreed baseline, thresholds, and reporting cadence.

Establish anomaly detection against it, and get the team to the point of trusting and using it.

Investigate flagged anomalies independently, with a root cause hypothesis and stated confidence.

Take ownership of:

Barcode and location verdict taxonomy and decision rules

The pattern library of known failure modes

Responsibilities include:

Version control

Closing documentation gaps

Adding edge-case guidance

Feeding changes back to the annotation vendor’s SOPs where the root cause is labeling

BS in Computer Science/Engineering, Electrical Engineering, or equivalent experience

2–5 years of experience in ML QA, annotation quality, data quality, or analytics at an AI/ML company

Hands-on experience with annotated ML datasets, including assessing label quality

Demonstrated experience finding, diagnosing, and reporting data anomalies to a technical audience

Able to get to a defensible answer from unfamiliar data without someone preparing it first

Understanding of data privacy and confidentiality requirements when working with customer operational data

Experience in warehouse automation, robotics, or computer vision applications

Experience with annotation platforms and quality tooling

Experience authoring annotation guidelines or standard operating procedures

Python

SQL, or an equivalent query language

Jira

We’re looking for someone who is:

Analytical, able to tell a real pattern from noise and to say honestly when the data will not settle a question

Detail-oriented — a verdict or a tracker that is subtly wrong is worse than none at all

Comfortable with ambiguity, and willing to present competing hypotheses rather than forcing a single answer

Persistent in chasing a cause across sites, programs, and builds

Resourceful and comfortable building processes and reports where they don’t yet exist

A strong written communicator — the output of this role is reports other people act on

Collaborative, respectful, and an excellent cross-functional partner

By the end of your first six months, you will have:

Established yourself as the primary owner of annotation quality analysis and root cause

A performance tracker the team relies on, running on a regular cadence

Closed the reporting cycle to same day, so issues are raised while still observable

Standardized the verdict taxonomy and decision rules for the categories you cover

Built a documented pattern library of known failure modes that others can use

Improved the annotation tool in at least one way that measurably reduces manual effort

Given ML and engineering a clearer picture of what error patterns mean for model accuracy and field performance

Created a repeatable method for standing up annotation quality on the next program