At Metamas, I deliver production AI services in Python, including Document AI, PDF-to-Markdown, anomaly detection, and an assessment platform. I’ve shipped these services as a founder or lead developer.
For utility work-order PDFs, I built an asynchronous extraction service that uses a Vision LLM to produce structured JSON. I expanded its coverage from 1 to 15 work-order types and more than 60 fields, with support for PDFs up to 94 pages.
I turned the AnomalyCLIP research model into a production service with a FastAPI backend and a React/TypeScript interface. On MVTec parts, statistical calibration raised defect recall from 15% to 80% and reduced false alarms from 27% to 0%.
I also built an Arabic assessment platform used to screen 200 university students at a Metamas hiring event, reducing manual grading by 93%. Across my work, I use FastAPI, Celery, and Redis to build AI services and asynchronous processing pipelines.
