Confidential industrial work
Production-Facing Computer Vision for Musculoskeletal Assessment.
As the sole AI engineer at a Germany-based startup, I developed an AI workflow spanning problem formulation, computer vision pipelines, validation, and product integration.
Problem
The product needed to turn user-captured mobile images from multiple views into structured anatomical measurements for musculoskeletal assessment. Inputs were imperfect, and model outputs needed to be checked for reliability before the AI workflow could support product-facing features.
Contribution
Zahra owned the AI workflow from problem formulation and experimentation through model development, validation, and product integration. She worked with physicians, orthopedists, and sports-medicine experts to translate clinical and product requirements into AI features, measurement logic, validation rules, and outputs supporting personalized corrective-exercise recommendations.
She worked with the FastAPI-based backend and Swagger/OpenAPI interfaces to test and validate AI API workflows, collaborating with the software engineer responsible for backend and server deployment.
System components
These are the principal components of the AI workflow, grouped conceptually without disclosing proprietary implementation order.
- Input viewsUser-captured mobile images from multiple body views.
- Visual analysisHuman detection, pose estimation, segmentation, and landmark localization/refinement.
- Quality controlView classification, confidence estimation, and anatomical plausibility checks for imperfect inputs.
- Geometric processingGeometry-based post-processing to produce structured anatomical measurements.
- Product integrationValidation of AI API workflows through the FastAPI/Swagger interface and coordination with the backend engineer.
Technologies used include Python, PyTorch, OpenCV, RTMPose, DensePose, SAM2, and MediaPipe. This is a conceptual component summary, not the company’s proprietary system architecture.
Engineering challenges & decisions
Imperfect mobile images: Developed human detection, view classification, and confidence checks to help assess whether inputs and predictions were suitable for downstream processing.
From predictions to useful measurements: Combined pose estimation, segmentation, landmark refinement, and geometric post-processing into structured anatomical outputs.
Reliability and domain constraints: Built anatomical plausibility checks and worked with physicians, orthopedists, and sports-medicine specialists to turn clinical needs into measurable rules and features.
Engineering trade-offs: Compared approaches with regard to accuracy, inference reliability, speed, and implementation complexity; tested AI workflows via the FastAPI-based backend and Swagger/OpenAPI interfaces alongside the backend engineer responsible for deployment.
Confidentiality
This work was part of a proprietary commercial application. Company datasets, source code, internal implementation details, and product outputs cannot be publicly shared. The case study describes my responsibilities and non-confidential technical approaches without claiming undisclosed performance figures.