

Lovro Iliassich
Verified Expert in Engineering
Machine Learning Developer
Rijeka, Croatia
Toptal member since February 20, 2016
Lovro is a machine learning engineer and data scientist who has worked in AI since starting his PhD in 2006 and brings full-stack software engineering experience. He builds custom machine learning and deep learning models as well as generative AI and multi-agent systems, choosing the right approach for each problem and delivering it end-to-end. Lovro's career spans research at INRIA and on a European Space Agency project, engineering at Amazon, and 20+ client engagements.
Portfolio
Experience
- Data Science - 15 years
- Machine Learning - 15 years
- Artificial Intelligence (AI) - 15 years
- Python - 10 years
- Predictive Modeling - 8 years
- Amazon Web Services (AWS) - 8 years
- Natural Language Processing (NLP) - 7 years
- Agentic AI - 3 years
Work Experience
Machine Learning Engineer | Data Scientist | Technical Screener
Toptal and Toptal Clients
- Delivered 20+ end-to-end AI engagements for clients across finance, healthcare, energy, education, media, and HR, usually as the sole machine learning engineer or the technical lead.
- Financial markets: extracted NLP features from earnings-call transcripts for a quantitative investment firm and architected a DeepSet neural network for equity return prediction, whose outputs fed the firm's final gradient-boosted model.
- * Multi-agent systems: architected agent-based assistants that decompose a request across specialized agents, including a flight assistant for pilots that coordinates several flight-planning agents behind a single conversational interface.
- Retrieval-augmented generation: designed RAG systems over large document corpora, including a clinical assistant that extracts values, reference ranges, and chart data from laboratory reports.
- * Natural language processing: built an LLM-based prototype for a global professional services firm that matches employees to internal projects by scoring their CVs against project descriptions.
- Computer vision: built a pipeline that detects diagnostic biostrips in photos and quantifies their color response with a trained neural network, and an OCR-based system that grades handwritten mathematics tests.
- * Healthcare and genomics: developed predictive models for early-stage cancer detection and chronic kidney disease risk from genomic data.
- * Forecasting and behavioral modeling: modeled EV charging station usage and user behavior, customer churn, lifetime value, and spending, and built product recommendation systems.
- Developed a convolutional neural network for environmental sound recognition and classification.
- Conducted over 500 technical interviews as a screener for Toptal's artificial intelligence and data science specializations.
AI Engineer | Solution Architect
Buildfast
- Built the generative AI and agent capabilities of a three-sided marketplace connecting homeowners, crew members, and the companies that hire crews for homebuilding and repair projects.
- Developed an AI video interviewing system that screens and qualifies crew members at scale, replacing manual first-round interviews.
- Built a conversational job search agent that matches crew members to open projects from natural-language queries.
- Generated quotes and project specifications automatically from meeting recordings, turning unstructured client conversations into structured, priced scopes of work.
- Architected the AI services and their integration into the existing marketplace platform.
AI Solutions Architect
Job.com
- Developed large language model applications for the recruiting industry, including conversational assistants for candidates and recruiters.
- Built retrieval-augmented generation pipelines over job listings and candidate documents.
- Developed candidate-to-job matching for the platform, scoring resumes against job listings with natural language processing models.
- Automated recruiting workflows with a queue-based AI service that parses CVs into structured candidate profiles and assesses interviews.
- Built smart candidate search and similar-candidate search over the candidate database.
Machine Learning Developer
Trust & Safety Laboratory Inc.
- Developed CatBoost and neural network models that detect misinformation in social media posts, used by some of the largest social networks to check posts on their own platforms.
- Implemented semantic matching by comparing posts against known misinformation using text embeddings and assessing each match across several axes.
- Developed and maintained the NLP pipeline, processing text from databases and reviews, and scraping web articles and social networks for misinformation content.
Software Development Engineer
Amazon
- Built back-end services in Java on AWS for the globalization of Try Before You Buy, Amazon's fashion retail program, extending an offering built for the US to European marketplaces.
- Adapted and deployed the program's services for individual European markets, each with its own locale, catalog, and marketplace rules.
- Developed the customer notification and transactional email services behind the program, including the templating used to localize messages per market.
Research Scholar
Drexel University
- Researched the mining and modeling of ranked and preference data with the Database Group, developing methods that reconstruct the preferences of a user population from partial input: rankings, partial orders, and pairwise comparisons.
- Implemented a Java library for representing, handling, and mining rank and preference data, used to run the group's experiments on real and synthetic preference datasets.
- Published peer-reviewed papers on the resulting approaches to modeling user preferences.
Research Engineer
INRIA
- Parallelized machine learning algorithms for high-performance computing, including SVM, affinity propagation, gradient descent, boosting, and neural networks.
- Optimized those algorithms in C/C++ for a large-memory (8 TB RAM) NUMA machine, tuning cache behavior, memory-block latency, and process-to-core assignment.
- Released apro (https://github.com/lovro-i/apro), a parallelized Java implementation of the affinity propagation clustering algorithm, as open source out of this work.
- Built a crawler and a category recommender system for Wikipedia on a semantic web project mining RDF and Wikidata.
Assistant Professor
Metropolitan University
- Taught undergraduate and graduate courses at the Faculty of Information Technology, including information systems design, web systems, web application development, and distributed systems.
- Developed the university's own information system, including the business process management workflow behind it.
- Authored and published the course material for the university's online courses.
Post-doc Researcher
University of Eastern Piedmont
- Built a real-time navigation system in C/C++ and OpenCV for a European Space Agency project, locating a Mars lander during descent from the live feed of its downward-facing camera.
- Detected and tracked distinctive surface features across the video stream, combining the visual position estimate with lidar and inertial measurement unit data.
- Modeled the Martian surface in Java 3D, simulating illumination at different times of day to generate imagery against which the vision system could be tested.
- Built a lander descent simulator in C/C++ and MATLAB.
PhD Student
University of Turin, Department of Computer Science
- Completed a doctorate in data mining and machine learning, working across text mining, sequential pattern mining, and complex network analysis.
- Performed text mining and document classification on public-sector documents for the regional government of Piedmont.
- Researched sequential pattern mining, including recognizing individual users from their keystroke patterns.
- Built usage models for a computational grid from log data, predicting network usage with data mining and complex network analysis.
Software Engineer
RCUB
- Designed and architected a real-time wide area network monitoring system, covering topology discovery, reporting, and intelligent agent assistance, deployed at large banks and telecommunications companies.
- Designed and implemented the hospital information system now used in about 75% of the hospitals in Serbia, covering medical records, patient scheduling, and clinical workflow, and led the team that built it.
- Worked across the full delivery cycle of these systems, from requirements interviews and UML specification through database design and back-end logic to web and stand-alone clients.
- Designed, architected, and implemented a government information system and a fleet management system.
Project History
Earnings-Call NLP for Equity Return Prediction
I built the ingestion and feature pipeline for earnings-call transcripts and applied transformer-based language models to extract sentiment, topic, and tone signals from both management and analyst speech. I then combined these text features with market, fundamental, and time-series inputs in a gradient-boosted ensemble that predicts equity returns.
The firm gained a pipeline that turns unstructured call transcripts into model-ready features alongside its conventional data.
Buildfast: Generative AI for a Homebuilding Marketplace
As an AI engineer and solution architect, I designed and built three AI capabilities and integrated them into the existing platform: an AI video interviewing system that screens and qualifies crew members, a conversational agent that matches crews to open projects from natural-language queries, and a pipeline that turns meeting recordings into structured, priced quotes and project specifications.
First-round crew interviews no longer need a person, and scopes of work come straight from client conversations.
EV Charging Demand and User Behavior Modeling
I analyzed historical charging-session data and modeled station-level usage patterns, temporal demand, and driver behavior.
The models supported capacity planning and utilization decisions across the network.
Job.com: AI Layer for a Recruiting Platform
https://www.job.com/As AI solutions architect, I designed and built that AI layer. At its core is a queue-based service that accepts different types of AI requests, such as parsing a CV into a structured candidate profile or assessing an interview, and processes them while throttling calls to the OpenAI API. On top of it I built smart candidate search, search for candidates similar to a given one, recommendations of the best-matching candidates for each job, and automated candidate interviews with AI assessment of the answers.
Many interviews no longer needed a recruiter, and recruiters got better-matched candidates for each job.
Multi-agent Flight Assistant for Pilots
I architected an AI assistant that handles both behind a single conversational interface. Specialized agents retrieve and interpret weather data, NOTAMs, and other flight-planning inputs, while a coordinating agent decomposes the pilot's request, dispatches it to the right agents, and assembles a grounded answer. For exam preparation, the assistant generates and evaluates practice questions grounded in aviation study material.
Pilots get one place to ask about a flight or an exam topic, with answers grounded in current data and source material.
Clinical Reporting Assistant
I built a generative AI assistant that ingests laboratory reports and extracts the clinically relevant information: measured values, their reference ranges, and data presented as charts rather than text. Using retrieval-augmented generation over medical guidelines and report templates, it recommends a patient-facing template and populates it with the extracted findings.
Physicians review and edit a draft instead of composing each report from scratch.
TaxBot: RAG Chatbot for a National Tax Service
I architected and implemented the whole system: a serverless back end on AWS (API Gateway, Lambda, DynamoDB), Qdrant for vector storage and retrieval over tax legislation and guidance, and a React front end. The OpenAI API generates context-aware answers grounded in the retrieved source documents.
I took the project from architecture to a deployed, operational service.
Misinformation Detection in Social Media Posts
I developed the NLP pipeline and models behind the detection service. Posts were represented as text embeddings and matched semantically against known misinformation; each match was assessed on several axes, and CatBoost and neural network models combined those signals to flag misinforming posts. I implemented detection for several major languages.
The social networks used the service to check posts on their own platforms.
Early Colorectal Cancer Detection from Genomic Data
I developed predictive models that classify early-stage colorectal cancer from genomic data, relying on careful feature selection, dimensionality reduction, and a validation strategy designed to prevent overfitting.
The approach kept performance estimates honest on data where naive models look deceptively good.
Satya: Constitution-Governed Conversational AI with a Self-Critique Loop
I designed the agent layer as a LangGraph loop. A drafting agent writes the reply under the persona prompt and the constitution. A critic agent checks the draft against the constitution and the safety rules and returns a structured verdict (approved, feedback, violated rules) instead of rewritten text. A rejected draft goes back to the drafting agent with the feedback until the critic approves it or a turn limit is reached. I built it first as a prototype, then as a self-contained package in the FastAPI back end, with each call going through the platform's LLM client with model failover, and the tokens of every call recorded so the user's prepaid wallet is charged for each.
Employee-to-project Matching for Internal Staffing
In early 2023, when large language models were still new, I helped build a prototype that scores how well an employee fits a project, working in a team of data scientists, engineers, and DevOps developers. Features extracted from CVs, text embeddings, and LLM-generated inputs, all drawn from the firm's HR database, were combined in a regression model that produced the match score.
The prototype was accepted, and the company that engaged me for early development continued to build it.
Automated Mathematics Test Assessment
I architected a pipeline that combines computer vision and OCR to locate answer regions on scanned test sheets, recognize handwritten mathematical content, and check it against the expected solutions.
The system scores tests automatically and flags ambiguous answers for human review, so teachers only look at the cases that need their judgment.
Member Churn and Lifetime Value Modeling for a Fitness Chain
Over a two-year engagement, I worked directly with the client's PostgreSQL database, writing SQL to assemble member-level datasets and developing models for member churn, customer lifetime value, and outlier detection.
The models gave the business a quantitative basis for retention and member-value decisions.
YieldHunter: Smart Glasses for Apple Harvest Monitoring
I built the computer vision, which runs on the glasses. I left out AR frameworks, which rely on flat surfaces and feature points that a leafy orchard lacks. An EfficientDet detector, running with TensorFlow Lite and MediaPipe, finds apples and the picker's hands, empty or holding an apple, in the live video. Each apple stays marked on the screen until it is picked, and a picking event, recognized across consecutive frames, increments the picked-apple count. A single model that also graded apples lost detection accuracy, so I split the work: the detector crops each apple and passes it to a classifier that grades it as first or second class, on criteria I worked out with a domain expert. For training data, I built a labeling tool that fruit growers without technical training could use.
Both models ran in real time on the glasses in a working prototype, with accuracy tuned against speed to keep the video smooth. The glasses fed a backend and dashboard, built by a colleague, that combined the counts from all pickers.
Automated Color Quantification for Diagnostic Biostrips
I built a pipeline that takes a JPEG photo and runs end to end with no human input. Classical computer vision techniques detect and isolate the test tube and the biostrip, and a neural network I trained compares the strip's colormap against a reference color spectrum.
Each strip is quantified numerically and graphically, replacing visual judgment with consistent, comparable readings.
Mountain Rescue Service of Serbia: Rescue Operations Information System
The system began in 2005 as a desktop application for writing and storing rescue reports. In 2020, I rewrote it completely: a serverless AWS back end (Lambda, API Gateway, DynamoDB, Cognito) and a React front end. It now covers rescuers and shifts, rescue actions, licensing, reports for ski resorts, and analysis, and exports reports to PDF.
The system is still growing and gains new features every season.
Sailweek: Fleet Operations and Crew Management System
http://sailweek.comSince 2022, I have architected and developed the company's information system, with a serverless AWS back end (Lambda, API Gateway, DynamoDB, Cognito) and a React front end. It began with the skipper and hostess workflow: staff management, boat assignments, and reporting. Each season adds features, including a boat database, docking arrangements, guest allocation to boats using optimization methods, and generative AI that helps skippers turn their own descriptions of boat damage into consistent, structured reports for the office.
The system keeps operations running smoothly for fleets of up to 60 boats and 500 guests, and continues to grow each season.
Heliant: Hospital Information System
https://heliant.rsAs an architect and team lead at the University of Belgrade Computer Center, I designed the system and led the team that built it. The work covered the whole cycle, from requirements interviews and UML specification through database design and back-end logic to the client applications. By 2006, it had grown from a single-hospital tool into a broader hospital information system.
After I left in 2006, it continued to grow and was largely rewritten. Today, it is the hospital information system of choice for about 75% of hospitals in Serbia.
Mandrago: Mobile Workforce Dispatching Platform
From 2013 to 2015, I co-founded Mandrago and served as its architect, designer, and developer, building the entire stack: infrastructure, back end, web front end, and Android app. The platform handled job assignment, location tracking, check-in and check-out, worker grouping, and bidding. It was designed to adapt to any business with a mobile workforce rather than to one client.
The startup did not find its market niche and was discontinued.
Education
PhD in Computer Science
University of Turin, Department of Informatics - Turin, Italy
Master of Science Degree in Computer Systems and Networks
University of Belgrade, School of Electrical Engineering - Belgrade, Serbia
Bachelor of Science Degree in Computer Science and Technology
University of Belgrade, School of Electrical Engineering - Belgrade, Serbia
Skills
Libraries/APIs
REST APIs, Scikit-learn, TensorFlow, PyTorch, Keras, OpenCV, XGBoost, Pandas, NumPy, CatBoost, OpenAI API, Matplotlib, LSTM, API Development, Google API, React, Claude API, Asyncio, Spark ML, PySpark, Google Maps API, Stripe API, Stripe, Google Vision API, Node.js
Tools
Adobe Photoshop, Algorithm Design, ChatGPT, Microsoft Excel, You Only Look Once (YOLO), Claude, Claude Code, GitHub, Amazon SageMaker, MATLAB, Amazon Cognito, Amazon Simple Notification Service (SNS), OpenAI Gym, Git, GIS
Languages
Python, Java, SQL, Python 3, HTML5, HTML, UML, C++, JavaScript, C, OWL, RDF
Frameworks
LangGraph, Android SDK, Serverless Framework, Agentic Frameworks, Streamlit, Ontology Framework, TensorFlow Lite, MediaPipe
Paradigms
RESTful Development, Parallel Computing, REST, Object-oriented Programming (OOP), Microservices, Automation, Functional Programming, ETL, Distributed Computing, High-performance Computing (HPC), Distributed Programming, Agile Software Development, UI Design, Machine-learned Ranking (MLR), Web UX Design, Unit Testing, Mathematical Optimization, Test-driven Development (TDD), Real-time Systems, Anomaly Detection, DevOps, Model Context Protocol (MCP)
Platforms
Jupyter Notebook, Web, Amazon Web Services (AWS), Amazon EC2, Android, Linux, Windows, Databricks, Google Cloud Platform (GCP), AWS Lambda, Docker, LangSmith
Storage
Database Modeling, Databases, Database Architecture, Data Integration, JSON, Data Validation, PostgreSQL, MySQL, NoSQL, Data Pipelines, Amazon DynamoDB, Google Cloud, Elasticsearch
Industry Expertise
Trading Systems, Healthcare
Other
Algorithms, Serverless, Predictive Modeling, Recurrent Neural Networks (RNNs), Unsupervised Learning, Clustering Algorithms, Clustering, Regression Modeling, Regression, Classification Algorithms, Classification, Deep Neural Networks (DNNs), Deep Learning, Convolutional Neural Networks (CNNs), Neural Networks, Computer Vision, Machine Learning, Data Mining, Data Modeling, Data Science, Scientific Computing, Data Visualization, Artificial Intelligence (AI), Natural Language Processing (NLP), Language Models, Web Development, Software Architecture, Back-end, Data Scientist, Predictive Analytics, Generative Pre-trained Transformers (GPT), Data Preprocessing, Feature Analysis, Software Engineering, Computer Science, LangChain, Large Language Models (LLMs), Chatbots, OpenAI, Prompt Engineering, K-means Clustering, AI Chatbots, Architecture, ChatGPT API, Forecasting, AI Model Training, Full-stack Development, AI Integration, Design, R&D, AI Consulting, AI Programming, Time Series Forecasting, FastAPI, AI Agents, Multi-agent Systems, Time Series Data, Agentic AI, Models, System Architecture, AI Enablement, AI Tools, Sequence Models, CSV, Agentic Workflow Design, Conversational AI, LLM Integration, Model Evaluation, Diagnostics, Model Validation, Validation, RAG Architecture, Backtesting, Demand Forecasting, AI Engineering, Front-end, Data, Prototyping, AI Workflow, Knowledge Bases, LLM Agents, AI Evaluation, AI Architecture, AI Inference, AI Infrastructure, Agentic RAG Systems, Datasets, Probability Theory, Modeling, Stock Market, Image Classification, Data Labeling, Document Automation, Bots, APIs, Full-stack, Data Architecture, Cloud, Recommendation Systems, Image Recognition, Sentiment Analysis, Geospatial Data, Cloud Platforms, Graphical Models, Image Processing, Data Analysis, Time Series, Time Series Analysis, Data Analytics, Data Reporting, Statistics, Visualization, University Teaching, Minimum Viable Product (MVP), BERT, Data Scraping, Distributed Systems, Natural Language Queries, Monitoring, Rankings, Spanish, Research, Graphical User Interface (GUI), Optimization, Retrieval-augmented Generation (RAG), Statistical Analysis, Generative Artificial Intelligence (GenAI), Vector Databases, Technical Leadership, Document Processing, Dashboards, Startups, Solution Architecture, Product Design, CTO, Consulting, Churn Analysis, RAG Systems, Anthropic, Computer Vision Algorithms, Images, Healthcare Software, Data Extraction, Integration, Reporting, Reports, Scripting, Training, Agentic AI Systems, AI-assisted Development, Software Development Lifecycle (SDLC), Technical Training, Embeddings from Language Models (ELMo), Multimodal Models, Sensor Data, Medical Software, API Integration, AI Agent Orchestration, AI Voice Agents, Evaluation, Workflow Automation, Financial Analysis, Prediction Markets, Financial Services, Semantic Search, AI Agent Evaluation, Maps, Principal Component Analysis (PCA), Access Control, Data Privacy, Observability, Sports, LLM Evaluation, Parsers, Logistics, Mixed-integer Linear Programming, Finance, Investment Strategy, Servers, Object Detection, LLM Reasoning, Data Anonymization, Data Masking, Data Protection, Trading, Data Engineering, OpenStreetMap, Stochastic Modeling, Web Scraping, Reinforcement Learning, Generative Adversarial Networks (GANs), Mapping, Optical Character Recognition (OCR), Amazon API Gateway, PDF, A/B Testing, Clinical Trials, Genomics, Computer Networking, Hugging Face, Deep Reinforcement Learning, Qdrant, Machine Learning Operations (MLOps), Health, Healthtech, Financial Markets, PDF Scraping, Equity, Equity Market Data, Customer Lifetime Value (CLV), Linear Optimization, Knowledge Graphs, AI DevOps, Supervised Fine-tuning (SFT), Tokenization, Large Language Model Operations (LLMOps), LLM Fine-tuning, Infrastructure as Code (IaC), MLflow, Privacy, Text-to-Speech (TTS), Speech Recognition, Ontologies, Geographic Information Systems, Statistical Methods, LoRa, GPU Computing, Edge AI, Human Resources (HR), Fintech, Algorithmic Trading, Algorithmic Trading Analysis
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
Share your needs
Choose your talent
Start your risk-free talent trial
Top talent is in high demand.
Start hiring