MANDRAGO
Applied AI · Machine Learning · Data Science · Software Engineering
Lovro Ilijašić

Lovro Ilijašić, Ph.D.

Also written Ilijasic, Iliassich, Iliasich

Machine learning engineer and computer scientist. I work at every level of AI, from first principles and classical models to deep learning and today's LLMs and agents. I use the one your problem needs.

Email me LinkedIn Toptal

01The whole stack

  1. 01
    Computer science

    Algorithms and the mathematics every model is built on.

  2. 02
    Statistics & classical ML

    Gradient boosting and statistical models with honest validation.

  3. 03
    Deep learning

    Neural networks for language, vision, and audio.

  4. 04
    LLMs & agents

    RAG and multi-agent systems built into real products.

  5. Software engineering

    The back end and product built around every level.

Anyone can call an LLM API now. The value is in knowing when an LLM is the right tool, when a well-validated classical model will do better, and when the problem needs a model trained on your own data.

02About

Mandrago is my practice. On larger projects I bring in a small team of collaborators I have worked with before. The name comes from a mobile workforce-dispatching platform I co-founded in 2013.

I have worked in AI since starting my Ph.D. in computer science in 2006. Before that, I spent five years as a software engineer. One of the systems I designed then grew into the hospital information system now used in about 75% of hospitals in Serbia.

Since then I have done research at Inria and on a European Space Agency project, engineering at Amazon, and more than twenty client engagements in finance, healthcare, energy, education, media, and HR. Because I came to AI from software engineering, I deliver the whole system: data, models, back end, and the product around them.

03Selected work

Clients are under NDA, so they are described by industry.

Deep learning Classical ML Quantitative finance

Equity return prediction from earnings calls

Problem
A quantitative investment firm wanted the signal in what is said on earnings calls: management's remarks, analysts' questions, and the answers they draw. Conventional market data does not contain it.
Approach
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.
Outcome
The firm now has a pipeline that turns raw call transcripts into model-ready features that sit alongside its market data.

Statistics & classical ML Healthcare R&D

Early cancer detection from genomic data

Problem
Colorectal cancer is far more treatable when caught early. Genomic data can help detect it, but it has many more features than samples, and models trained on it overfit easily.
Approach
I built classifiers for early-stage colorectal cancer from genomic data, using careful feature selection and dimensionality reduction. I designed the validation strategy specifically to catch overfitting.
Outcome
The client got performance estimates it could trust, on data where naive models look deceptively good.

LLMs & agents Homebuilding marketplace

Generative AI across a three-sided marketplace

Problem
A marketplace connecting homeowners, crews, and contractors screened crews, matched them to projects, and turned client meetings into quotes by hand.
Approach
As AI engineer and solutions architect, I designed AI capabilities and built them into the existing platform. These included AI video interviews that screen and qualify crew members, a conversational agent that matches crews to open projects from plain-language requests, and a pipeline that turns meeting recordings into structured, priced quotes.
Outcome
First-round crew interviews no longer need a person, and quotes and scopes of work come straight from recorded client meetings.

04Start small

2 weeks · Fixed scope

ML/AI feasibility sprint

For when you suspect machine learning or AI could solve a problem but are not yet sure how, or whether it is worth it. Better to find out in two weeks than halfway through a build.

Ask about a sprint

What you get

  1. An assessment of the problem and of the data you already have.
  2. The simplest baseline that could work, so every later result has something to beat.
  3. A working prototype of the most promising approach: classical, deep learning, or LLM, whichever fits.
  4. A written go/no-go with the proposed architecture and an estimate of build and running costs.

05Contact

lovro@mandrago.com
LinkedIn Toptal For agencies