Davide Zagami
About Me
I’m a Deep Learning engineer with 5+ years experience building and shipping complex data-driven solutions.
I also have a passion for research in Artificial Intelligence and Decision Theory, having previously researched problems in these fields and presented one of my publications at the biggest AI conference in the world (IJCAI).
Knowledge areas: Object Detection & Segmentation · Image Classification & Recognition · Named Entity Recognition · Word Embeddings · Document Understanding · Data Annotation Lead · CICD · Orchestration
Technologies: PyTorch · HuggingFace · Tesseract · spaCy · Flair · scikit-learn · pandas · GitLab · Kubernetes
Projects
Konfuzio SDK
I was responsible for the regular release of the open-source Konfuzio SDK package, a Python library to extract, classify and split documents using a combination of Tesseract, Pytorch, Keras, Transformers, Tokenizers and sklearn.
FUNSD+
I spearheaded the development of FUNSD+, an enhanced and expanded iteration of the prominent FUNSD dataset, by selecting documents from RVL-CDIP and coordinating with a team of 3 annotators.
Visit https://app.konfuzio.com/d/303962/ (just deny credentials when asked)
Hora Video Analytic System
HoraVision is a start-up I co-founded that developed a Computer Vision cloud platform. Among many other Deep Learning models, I developed and deployed a face and person recognition model in Python using neural network architectures including Yolo and EfficientNet, as well as various features such as gender, age, head and body pose, gaze estimation, landmarks, mask usage and clothing. These models can detect more faces quicker than Google Cloud, Amazon Rekognition, and Azure Vision.
Find the SaaS at app.horavision.ai and an overview on GitHub.
Tools for drawing candlestick charts and several trading indicators, together with an API for backtesting and deploying strategies. Example: Live buy and sell during new Binance coin listings (7% expected profit per listing).
See GitHub for charting/backtesting tools, a coin listing trading bot, and other tools.
RAISE
I was the head of a team of three creating an online course about Inverse Reinforcement Learning and I wrote code to analyze the data of a Randomized Controlled Trial involving 100+ students learning from my lessons.
Categorizing Wireheading in Partially Embedded Agents
Mathematical and experimental exploration on wireheading, the behavior of corrupting the internal structure of an AI agent in order to achieve maximal reward without solving the designer’s goal. IJCAI paper here.
Davide Zagami
About Me
I’m a Deep Learning engineer with 5+ years experience building and shipping complex data-driven solutions.
I also have a passion for research in Artificial Intelligence and Decision Theory, having previously researched problems in these fields and presented one of my publications at the biggest AI conference in the world (IJCAI).
Knowledge areas: Object Detection & Segmentation · Image Classification & Recognition · Named Entity Recognition · Word Embeddings · Document Understanding · Data Annotation Lead · CICD · Orchestration
Technologies: PyTorch · HuggingFace · Tesseract · spaCy · Flair · scikit-learn · pandas · GitLab · Kubernetes
Projects
Konfuzio SDK
I was responsible for the regular release of the open-source Konfuzio SDK package, a Python library to extract, classify and split documents using a combination of Tesseract, Pytorch, Keras, Transformers, Tokenizers and sklearn.
FUNSD+
I spearheaded the development of FUNSD+, an enhanced and expanded iteration of the prominent FUNSD dataset, by selecting documents from RVL-CDIP and coordinating with a team of 3 annotators.
Visit https://app.konfuzio.com/d/303962/ (just deny credentials when asked)
Hora Video Analytic System
HoraVision is a start-up I co-founded that developed a Computer Vision cloud platform. Among many other Deep Learning models, I developed and deployed a face and person recognition model in Python using neural network architectures including Yolo and EfficientNet, as well as various features such as gender, age, head and body pose, gaze estimation, landmarks, mask usage and clothing. These models can detect more faces quicker than Google Cloud, Amazon Rekognition, and Azure Vision.
Find the SaaS at app.horavision.ai and an overview on GitHub.
Cryptocurrency trading tools
Tools for drawing candlestick charts and several trading indicators, together with an API for backtesting and deploying strategies. Example: Live buy and sell during new Binance coin listings (7% expected profit per listing).
See GitHub for charting/backtesting tools, a coin listing trading bot, and other tools.
RAISE
I was the head of a team of three creating an online course about Inverse Reinforcement Learning and I wrote code to analyze the data of a Randomized Controlled Trial involving 100+ students learning from my lessons.
Categorizing Wireheading in Partially Embedded Agents
Mathematical and experimental exploration on wireheading, the behavior of corrupting the internal structure of an AI agent in order to achieve maximal reward without solving the designer’s goal. IJCAI paper here.