Internal review & operations tool
VerifAI
Project owner · Engineering team lead
Developed as Software Engineer; continued ownership as Lead Software Engineer
I own VerifAI and lead its engineering team. My work covers system design and implementation for reviewing, correcting, and auditing AI card identification and image quality, with controls for managing review operations. The card dataset I work with contains more than 2.5 million entries.
THE BRIEF
What needed to be built.
Reviewers need to check AI card identification, assess image quality, correct errors, and audit the results. The team also needs controls to manage these review operations.
MY CONTRIBUTIONS
My contribution.
Review workflows
- Built review, inspection, assignment, correction, and audit workflows with React, FastAPI, PostgreSQL, and MongoDB.
- Centralized roles and permissions, enforcing access at API and interface boundaries.
- Saved reviewer progress in PostgreSQL across reloads and order changes.
Data & reliability
- Made queue ingestion safe to retry; replaced long-lived database sessions with scoped transactions and rollback handling.
- Improved indexing, filtering, pagination, caching, and polling; load-tested worker and EC2 configurations.
- Develop recovery safeguards with row locks, worker ownership checks, and interrupted-job recovery.
Operations & maintenance
- Maintain inherited Postcard sync, SiteCache, and image-transformation services; investigate cross-service failures.
- Added automated checks and runbooks for permissions, queue transitions, database behavior, and recovery.
WHY THIS APPROACH
Treat reviewer progress and queue state as durable application data. Persist work before navigation, make repeated ingestion safe to replay, and keep database sessions short. Explicit permissions and ownership checks establish who can change each workflow.
THE RESULT
Human review backed by access controls and recoverable workflows.
Built the review and assignment platform during my Software Engineer role and continue to lead its development. The platform connects human review with persisted progress, access controls, and operational tools for investigating failures. The card dataset contains more than 2.5 million entries.