Undergraduate thesis · Clinical NER in Spanish

Corpus annotated from scratch, four modeling approaches compared. 0.8767 strict F1 on a held-out test set, dropping to 0.5636 against independent expert annotations — a gap in annotation criteria, reported as the study's main limitation. Served through a FastAPI microservices backend exposing five selectable models, each isolated in its own container.

Neural architectures behind the thesis

The BiLSTM+CRF baseline the thesis transformer had to beat. Custom Keras training loop over the CRF log-likelihood instead of a per-token loss, plus a Dice objective for the heavy O-tag imbalance. Pretrained Spanish Word2Vec embeddings, BIO tagging, grid search over batch size and epochs.

Pragma Power Up · E-commerce backend

Four REST microservices — users, catalog, cart, checkout — built on hexagonal architecture, with JWT auth and inter-service calls over OpenFeign that propagate the caller's token. 53 unit test classes.

Multi-site tool & fleet management

Inventory, loans and inter-site transfers for workshops with multiple locations. Lead contributor on both ends — 138 backend classes and 10 versioned Flyway migrations, plus the React frontend.

3D adventure game in the browser — play it →

Four levels, physics-driven movement, spell and mana systems, real-time multiplayer over Socket.IO. Full UX cycle behind it: interviews, affinity mapping, protopersonas and per-level usability testing.

Hackathon · Shared LLM agent session — try it →

An agent session a whole team watches live. Every instruction is attributed, conflicting requests open a vote that halts the agent, and control passes between drivers without losing context.

Neural Networks coursework · Four architectures, one task

MLP, RNN, LSTM and a fine-tuned RoBERTa predicting review ratings, each under three treatments of class imbalance. The result worth reading is a negative one: loss weighting failed to remove majority-class bias in every configuration, undersampling traded accuracy for fairness, and scale erased the tradeoff. Bayesian search over 30 trials; the full Yelp dump streamed as NDJSON into Parquet shards to fit Colab. Ships with the 37-page report.

NLP coursework · Two routes to the same problem

A BiLSTM-CRF with in-domain Word2Vec on the public CodiEsp corpus, against BETO and XLM-RoBERTa fine-tuned with LoRA adapters over ten oncology entity types including Gleason grading and TNM staging. Swept across epochs and batch sizes, scored with seqeval at entity level, and reported as macro F1 0.905–0.933 rather than the weighted average the O tag inflates. Adapter published to the Hub.