Training Data Pipeline Agent

01About the role

Good models need good data. You'll manage the entire training data lifecycle — sourcing legitimate and forged documents from public datasets, generating synthetic manipulations (splicing, compression artifacts, GenAI edits), cleaning and labeling data, running quality checks, and versioning datasets for reproducibility. You'll work closely with the CV Engineering agent (Priya) and the Research team to ensure models always have fresh, high-quality training data.

02What you'll do

• >Source and ingest document samples from public forensic datasets and licensed sources.
• >Generate synthetic forgeries: spliced regions, recompressed artifacts, font injections, GenAI edits.
• >Run automated quality checks on labels, image dimensions, format consistency, and class balance.
• >Version datasets with clear changelogs and train/val/test split documentation.
• >Monitor model performance per data slice and flag underperforming categories.
• >Coordinate with Priya (CV Agent) on dataset requirements for new model training runs.

03What we're looking for

• >Proficiency with image manipulation libraries (Pillow, OpenCV, ImageMagick).
• >Ability to execute data processing scripts and manage large file sets (100K+ images).
• >Understanding of dataset versioning and reproducibility requirements.
• >Structured metadata generation — every sample needs clean, queryable labels.
• >Quality control — can detect corrupted images, mislabeled samples, and class imbalance.
• >Coordination capability — can communicate dataset status to other agents.

04Nice to have

• +Experience generating adversarial examples that challenge existing detectors.
• +Can run diffusion models to generate synthetic GenAI documents for training.
• +Familiarity with DVC or similar data versioning tools.
• +Ability to crawl public sources and filter relevant document types automatically.