Everyone calls me Medi. I'm an NLP researcher and PhD student in computer science at University Politehnica of Bucharest. My research focuses on low-resource languages, especially Romanian. I also work as a data scientist at Endava.
NLP researcher, PhD student in computer science, data scientist, Python teacher and software engineer.
My name is Ana-Maria Luisa Mocanu, but everyone calls me Medi. I live in Bucharest, Romania.
I'm a PhD student in Computer Science at University Politehnica of Bucharest, where I also earned a Master's in Artificial Intelligence and a Bachelor's in Mechanical Engineering. My research is in natural language processing for low-resource languages, especially Romanian. I build datasets and annotation tools, work on aspect-based sentiment analysis, and distill what large language models know into small models that work well for Romanian. Alongside that, I've worked on humor retrieval and multimodal learning.
Professionally, I'm a Data Scientist at Endava, applying machine learning and NLP to data-driven problems, from fraud and risk models on card-payment data to LLM solutions and AI agents on Google Cloud, after four and a half years there as a software engineer building fintech services. Beyond research, I teach Python to bachelor's students.
I apply machine learning and NLP to data-driven problems. The role builds directly on my research and on the software-engineering background from my fintech projects.
I build ML models on card-payment data, such as fraud and risk scoring and authorization-rate optimization. I explore, clean and analyse transaction data in Python and SQL, engineer features and present findings to clients, and I build LLM-based solutions and deploy AI agents on Google Cloud (Gemini Enterprise).
I built and maintained fintech applications from scratch in Java and Spring Boot across several projects, one of them nominated among Endava's most successful projects. I grew from feature work into owning test automation, DevOps tasks, deployments, Kong gateway configuration and release preparation, working closely with clients and product owners.
Clients included a top-3 UK bank, a premium German carmaker's payments system and a European e-invoicing platform, delivered in service-oriented and microservice architectures within an Agile team.
I teach beginner-level Python to bachelor's students in Google Colaboratory, guiding them through lab exercises, projects and tests, and helping them build practical programming skills.
I design the lab exercises, homework and tests, I've received 12 positive student reviews for making programming concepts easy to understand, and I co-coordinate 2 bachelor's and 2 master's students on their thesis projects. I also built Scholar Rank, the auto-grading platform we use for live lab exams.
Research in artificial intelligence: designing models and applying machine learning to complex natural language problems.
Thesis: Methods and Resources for Low-Resource Languages. Coordinators: Prof. Ciprian-Octavian Truică and Prof. Elena-Simona Apostol; guidance committee: Prof. Liviu P. Dinu and Prof. Mihai Dascălu.
Advanced studies in artificial intelligence, machine learning and NLP.
Thesis: Aspect-Based Sentiment Analysis for Romanian Product Reviews. Coordinators: Prof. Ciprian-Octavian Truică and Prof. Elena-Simona Apostol.
A foundation in engineering and technical problem solving.
Thesis: Study on Remote Monitoring Possibilities for a Refrigeration System. Coordinators: Prof. Valentin Apostol and Prof. Horaţiu Lucian Pop.
IROH ranked 1st on JOKER Task 1 English (0.6347 MAP), and our multimodal sexism detection system ranked 29th of 114 on EXIST Subtask 2.2. More under Publications.
Presented ACAT, our collaborative annotation platform. More under Publications.
I helped coordinate coastal conservation and community-led waste management, and ran workshops on ecological responsibility and the circular economy.
I organised workshops and charity events for social welfare and youth development, which taught me a lot about communication and handling logistics on tight schedules.
I worked on large-scale charity events for local communities, learning to take initiative under pressure.
1st place, 0.6347 MAP
29th of 114, Subtask 2.2 (soft evaluation), top 26%
1st place, CLEF 2026 JOKER
BM25 + dense retrieval / cross-encoder reranker / LoRA and QLoRA LLM judges
A three-stage pipeline combining hybrid sparse-dense retrieval, a fine-tuned cross-encoder reranker and an ensemble of LoRA-adapted LLM judges trained on distilled rationales. It ranked 1st on the official JOKER 2026 Task 1 English leaderboard with a MAP of 0.6347.
Top 29 of 114, CLEF 2026 EXIST
CLIP + XLM-RoBERTa / FiLM conditioning / cross-attention
A human-centered model that fuses meme text, a generated visual description, the image, annotator demographics and physiological signals. Cross-attention grounds the description in the image, and FiLM conditions the prediction on who is annotating.
Master's thesis
GPT-4.1 / Claude Haiku / RoBERT / BERT Romanian cased
I created the Romanian Review (RoR) dataset: 30,000+ e-commerce reviews annotated across 14 product aspects, combining manual labels with LLM annotation. Distilled into ~100M-parameter Romanian transformers, the students reached F1 0.787–0.788, on par with their GPT-4.1 teacher.
Flask / PostgreSQL / Docker
A web platform for annotating four aspect-based sentiment tasks of increasing granularity. It computes inter-annotator agreement (Cohen's and Fleiss' kappa) at export time and exports to CSV, JSON and XML. It's currently used by 400+ users.
scikit-learn / LightGBM / PyTorch LSTM / FastAPI / Strands Agents
I cleaned a messy 41.6K-row card-transaction dataset and split it by time to prevent leakage, then compared Random Forest + SMOTE, LightGBM and an LSTM at a 1.4% fraud rate. LightGBM reached PR-AUC 0.65 on the held-out month (73% precision at 61% recall). The model is served with FastAPI, and an AI agent calls it as a tool and remembers each conversation.
PyTorch / TensorFlow / Django / Twitter API
We trained 7 CNN and transformer models on 5 hate-speech datasets, combined them with soft and hard voting, and packaged the result into a Django web app connected to the Twitter API.
Logistic regression / BERT / BERTweet Large
A Kaggle competition on predicting which tweets are about real disasters. My best submission reached 0.84554 accuracy with a fine-tuned BERTweet Large model.
Flask / PostgreSQL / Docker / Monaco editor
An exam platform where students solve Python problems in a browser IDE and their code is graded automatically in sandboxed Docker containers. It has a professor dashboard, randomized problem pools and anti-cheat measures, and it's used for live lab exams.
Django / MySQL / JavaScript
A full-stack Django site for a university department, with an admin-friendly CMS so academic staff can edit pages and manage content without technical help.
Java / Android Studio / Room
An app that helps bunny owners plan healthy meals and track daily diets, to prevent the health problems that come from the wrong food.
Django / SQLite / Spotify Web API
A small social network built on the Spotify Web API, with full CRUD for posts and a backend designed for future recommendations based on each user's music taste.