Romain Florentz

A coding agent built a deterministic German verb conjugator, driven by a verification loop against two morphology gold standards: from 62.8% to 99.2% agreement.

Comparing spaCy and LLM structured generation for named entity recognition on Dune, building a social graph of characters from the novel's text.

Training a small GPT on x + y = z equations to measure exactly how well a transformer learns addition, with an objective score instead of a vibe check.

An interactive dashboard comparing the dimensions of published transformer LLMs: parameter counts, feed-forward ratios, embeddings and attention heads.

Generating a complete set of connected, symmetry-distinct glyphs on a 3x3 grid using equivalence classes under the dihedral group and an incremental search algorithm.

Lessons from improving LeNet training on Armenian script: dataset generation, augmentation, TensorBoard monitoring, and a Hugging Face Space demo.

How diffusion models connect to a 40-year-old result on time reversal of diffusion processes, with a from-scratch implementation on a Gaussian mixture.

Building a character recognition engine for Armenian script with PyTorch: generating a dataset from fonts, training LeNet-5, and evaluating across typefaces.

Running Tesseract OCR on 1920s Armenian newspaper scans: image enhancement, measuring character error rate, and where the engine struggles.