Jovan Spasojevic

Created affective-computing AI models that are combining both expert knowledge and their intuitions, to calculate the quality score of complex decisions. Created an NLP classification algorithm for legal documents corpora based of the NLTK library, constructed using mixed feature-extraction techniques: POS-Tagging, noun-phrase extraction, collocations and NER (named entity recognition), followed by Tf/Idf, feature reduction and finally the classification with Passive-Aggressive, scalable classifier.

Created a critical part of a tax-fraud detection system was based of natural languages rules enabling decision makers and specialists to manage a tax fraud knowledge base. The stream-based reasoner allows discovering fraudulent activities in the stream of 5 million invoices per day.

Technologies: Apache Jena, Simple Knowledge Organization System (SKOS), SPARQL, Semantic Web Rule Language (SWRL), RDF, OWL, Business Process Modeling Notation (BPMN), TensorFlow, NumPy, Scikit-learn, Natural Language Toolkit (NLTK), R, Python, Minimum Viable Product (MVP), Artificial Intelligence (AI), Generative Artificial Intelligence (GenAI)

Architected Machine Learning and Deep Learning pipelines that extracted metadata from 10,000+ hours of media content, cutting manual tagging effort by 80%. Launched 6 proofs of concept across Computer Vision, Natural Language Processing, and Multimodal Learning for automatic content tagging; promoted 3 to production. Tuned multimodal activity and event detection classifiers across video, audio, and text modalities, achieving 87% F1 on internal benchmarks.

Pioneered an unsupervised entity extraction and entity linking algorithm for a sports news platform, earning 2 patents. Technologies: Python, PyTorch, Machine Learning, Deep Learning, Data Science, Amazon Web Services (AWS), Relational Databases, Natural Language Processing (NLP), Generative AI (GPT), Statistics, Multimodal Learning, Computer Vision

Advised executive stakeholders on AI opportunities and shipped a working MVP within 12 weeks, validating the proposed solution on production client data. Architected the AI pipeline from data acquisition to prediction for life-science systematic-review workflows, processing 10,000+ regulatory documents per run. Guided the client infrastructure team in standing up a production MLOps pipeline (model registry, CI/CD, monitoring) that cut deployment lead time by 60%.

Productionized question-answering systems for clinical and regulatory documents that returned answers in under 2 seconds at 91% retrieval precision. Constructed an automated inclusion/exclusion classification pipeline for systematic reviews, reducing reviewer workload by 70%. Directed a 4-person team through agile weekly ceremonies and informed executive go-to-market positioning of the AI product within the life-sciences sector.

Technologies: Machine Learning, Artificial Intelligence (AI), Generative AI (GPT), Natural Language Processing (NLP), Python 3, PyTorch, Language Models, Docker, MLOps

Stress-tested and pre-flighted 10+ Shopify client storefronts ahead of Black Friday, sustaining 100% uptime through a 4x traffic peak. Programmed a reusable headless eCommerce application on Express.js and Vue.js consuming the Shopify Storefront API, deployable across any Shopify store in under 1 week.

Released a custom Shopify app extending product pages with administrator-managed supplementary content, adopted by 15+ merchant stores. Technologies: Shopify Plus, HTML, CSS, JavaScript, Express.js, Vue.js, Headless Commerce, Storefront API