Open to new opportunities
I'm Andrea, a computer vision engineer. My core is industrial quality control, metrology and real-time detection, tracking and pose — and I build the whole product around it: native iOS apps, backends and data pipelines, from edge devices to the cloud.
Expertise
Automated visual inspection and defect detection, plus classical computer vision for precise measurement: camera calibration, sub-pixel edges and dimensional checks.
Custom-trained detectors and multi-object tracking for people, vehicles, products or anything that moves in a video.
Keypoints, actions and temporal events — for sport, ergonomics, safety or retail.
Models optimized for Jetson, Raspberry Pi and on-premise boxes: quantization, ONNX/TensorRT export, real-time pipelines.
From RTSP cameras to dashboards: ingestion, inference, storage, APIs and live monitoring.
Native SwiftUI apps with models running privately on the phone, and the backend behind them: Postgres/Supabase, FastAPI services and React dashboards.
Selected work
Each project solved a hard constraint — privacy, latency, limited hardware or messy real-world footage.
A plug-and-play edge box that turns a store's existing cameras into anonymous people trajectories on the floor plan — footfall, dwell, heatmaps and queue KPIs. Privacy by architecture: no face recognition, no frames leave the store, only aggregated events go out.
Detection → tracking → pose → temporal event models. The system learns reusable atomic actions (reach shelf, pick object, hand to pocket) and reasons over them to produce an explainable risk score — instead of a black-box 'shoplifting' classifier.
Finds made shots in amateur game footage and exports clips plus a highlight montage. Ball detection with camera-motion-compensated tracking, auditable rule-based classification, and an evaluation harness against manual annotations. Runs fully offline on a laptop.
An iPhone app that analyzes side-on videos of indoor rowing: pose estimation on the phone, automatic segmentation of each stroke into catch, drive, finish and recovery, knee, hip, elbow and trunk angles, and stroke-to-stroke consistency on an annotated replay. Backed by a reproducible research harness that benchmarks pose models on Mac and on a physical iPhone before choosing one.
Detection and recognition models ported from Python/ONNX to a native iPhone app: find every photo of a person without a single image leaving the phone.
Graph-based inference runtime with sessions, runs and pluggable 'applications' over cameras, RTSP streams or uploads. Decoupled workers, retention policies, Prometheus metrics and a React control room.
Segmentation, rider keypoints and motion trails, graded frame by frame against the technical criteria of the Italian equestrian federation manuals.
Order PDFs parsed and matched against the ERP by a local LLM, proposed to a human for approval, then written back with a full audit trail.
A native iOS app that shows tonight's events on a map and lets you see where the people you follow are going. Clean modular architecture, Supabase backend with geospatial queries and privacy rules enforced by row-level security.
About
I'm a computer vision engineer with more than three years of experience building vision applications for industrial and commercial clients — from designing and training the models all the way to deploying them on edge devices. I specialize in automated quality control and classical computer vision for metrology, alongside deep learning for detection, tracking and pose in real-time video.
I like hard constraints: tight measurement tolerances, footage that cannot leave a store, a phone that has to do everything offline. That's where careful engineering beats a bigger GPU.
Measured, not promised
Every claim comes with an evaluation on real data.
Privacy by architecture
On-device and on-premise whenever possible.
Built to last
Tested, documented, reproducible.