Property-Tech Vision AI Platform (Confidential)
A production computer vision and VLM pipeline that turns property photos into structured, customer-ready outputs.
Computer VisionVLMFastAPIMLOpsSaaS
A production computer vision and VLM pipeline for a property-tech SaaS platform serving US customers: property photos go in, and the system detects, segments, and classifies their contents through a multi-stage pipeline, turning raw image uploads into structured, customer-ready output end to end. I own the ML pipeline architecture from raw image input through to automated output, shipped via FastAPI and Docker with MLflow model versioning and GitHub Actions CI/CD. Production details are confidential, happy to walk through the architecture in interviews.
What I built
- Designed and built the backend service layer, ingestion, processing, structured output generation, and admin review, across a routers, services, and integrations architecture.
- Built a custom session and authentication layer with role-based access control.
- Built a fully programmatic report generator producing a structured, audit-ready output document per customer submission.
- Built a human-in-the-loop admin correction workflow as the foundation for ongoing pipeline refinement.
- Own the deployment and MLOps roadmap for the pipeline's classification stage, which calls a hosted vision-language model rather than a self-hosted model.
Tech stack
Computer VisionVLMFastAPIMLOpsSaaS