About SelfInspection SelfInspection is building a Vehicle Condition Intelligence Platform: the system of record for vehicle condition across a vehicle's entire lifecycle. Our AI turns fragmented inspection and handoff data into a trusted, continuously updated source of truth. Rental operators, fleets, dealerships, finance providers, insurers, transportation companies, automotive marketplaces and mobility platforms use it to manage risk, protect assets and reduce disputes. The role You'll own the data behind our vehicle-condition AI: every image, video, label and dataset our models learn from. You'll build the pipelines that turn production inspection media into curated, labeled, versioned datasets, and you'll run our labeling operation in CVAT. This is a senior, hands-on role for someone who has run computer vision data at scale before. You'll set the standards for data quality, labeling and governance, and work daily with our CV and ML engineers. Full-time, remote.

What you'll do
• Own the CV data platform. Pipelines that move production inspection images, video and metadata into a searchable, versioned dataset store on AWS.
• Run labeling in CVAT. Administer CVAT end to end: projects, label schemas, tasks, users, review stages and exports, automated through its API and SDK.
• Make labels trustworthy. Annotation guidelines for our damage taxonomy, reviewer workflows, gold tasks, inter-annotator agreement and regular audits.
• Speed up labeling. Model-assisted pre-annotation with SAM-style models and our own detectors, plus active learning that sends hard cases to annotators first.
• Manage annotators. Onboard, train and QA in-house and vendor annotation teams,



and track throughput, cost and quality per batch.
• Curate and version datasets. Train, validation and golden splits, deduplication, class balance, and lineage from every model back to its source inspections.
• Govern the data. Access control, retention, and handling of license plates, faces, VINs and location data that hold up to enterprise security reviews.
• Feed training and evaluation. Clean exports (COCO, YOLO, CVAT XML) and dataset releases wired into training pipelines and MLflow. What you bring
• 8+ years in data engineering, including 3+ years building and managing datasets for computer vision or ML.
• Hands-on CVAT at scale: administration, label schema design, review workflows, and automation through its API and SDK.
• Strong Python and SQL, and production pipelines in Airflow, Dagster or Prefect.
• AWS data stack in production: S3, IAM, ECS or Batch, Lambda, Athena or Glue, and PostgreSQL.
• Large media datasets: millions of images and hours of video, with deduplication, metadata, and fast search and retrieval.
• Label quality management: guidelines, agreement metrics, gold sets and audits, plus leading in-house or vendor annotation teams.
• Dataset versioning and lineage with DVC, lakeFS, FiftyOne or similar.




• Privacy and security for image data: PII detection and redaction, access control and retention policies.
• Fluent professional English, spoken and written. Our team works in English every day.
• Clear writing, and the confidence to set data standards and push back on bad data.
• BS/MS in Computer Science, Data Science, Electrical Engineering or a related field, or equivalent practical experience. Nice to have
• Automotive damage, inspection, insurance-claims or other visual-inspection datasets.
• Model-assisted labeling: SAM or SAM 2, CVAT serverless (Nuclio) functions, active learning loops.
• Embedding-based curation: CLIP-style embeddings, vector search, near-duplicate and outlier detection.
• Video annotation: frame sampling, keyframe extraction, and object tracks in CVAT.
• Self-hosting CVAT on AWS with Docker or Kubernetes, managed with Terraform.
• Working under ISO 27001 or SOC 2 controls.
• MLflow, and a working understanding of how CV models are trained and evaluated. What success looks like in 6 months
• Every production inspection's images and video land in a searchable, versioned store with lineage to model releases.
• CVAT runs as a managed service, with schemas, review stages and API automation, and throughput and quality reported weekly.
• Labeling guidelines cover the full damage taxonomy, with agreement measured and gold tasks in every batch.
• Model-assisted pre-labeling has measurably cut labeling time per image.
• Golden evaluation sets for damage, tires and image quality are versioned, owned and used by every model release.

📌 Senior Data Engineer (México)
🏢 Self Inspection
📍 México