Heonir Feliz

Neuro-Symbolic Legal Knowledge Graphs

with Bloomberg

Built an AI system that converts unstructured legal narratives into machine-reasonable knowledge graphs over 10k+ extracted events and relationships, combining LLMs, LegalBERT, and graph reasoning to deliver a +613% improvement over TF-IDF baselines via explicit temporal and causal reasoning.

AI Poker Agent

Reduced the game state from 3 trillion to under 1,000 representations per play (99.7% reduction) and trained over 10,000+ simulations, hitting an 80% win rate under a 95 ms decision-time limit.

Large-Scale Semantic Retrieval on Jeopardy! Data

Built an unsupervised pipeline clustering 216,000+ Jeopardy! Q&A pairs into coherent topics with text embeddings, tuned via elbow method and silhouette analysis, then layered a semantic-similarity retrieval system that classifies new questions and surfaces related clues.

LeetCode Solutions

— solutions and counting

An always-growing collection of my LeetCode problem solutions, each in its own folder with a write-up of the approach. The count above is pulled live from GitHub on every visit.