MISSION DOSSIERS
Portfolio
Case studies with the full story: the problem, the architecture, the experiments that failed, and the numbers that survived. For the raw repositories, see the live GitHub projects on the home page.
-
LabML - ML in Your Browser: a Complete Machine Learning Lab With No Backend
LabML is a full tabular ML platform - data quality, training, evaluation, explanation, model reuse - that runs entirely in the browser: hand-written, seeded algorithms in Web Workers, no backend, no accounts, no uploads. Live at app.dominicdapice.com, with an ML Lab, a Data Studio with in-browser analytical SQL, and an AI playground with on-device vision and a local language model.
-
Titanic on Kaggle - When Better Local Validation Doesn't Mean a Better Score
A solo deep-dive into the classic Kaggle Titanic competition, pushed beyond the usual notebook: tested feature engineering, ticket-grouped validation, four model families compared - and a Gradient Boosting that scored 0.8440 locally but 0.75358 on Kaggle. Why I stopped at an honest 0.78947 instead of chasing the leaderboard.
-
Mila Hackathon - Detecting a Crisis in Two Languages: A Hybrid mmBERT × Mistral Guardrail
Inside Mila's first AI safety hackathon (March 2026): how our team built a bilingual mental-health risk classifier - fine-tuned mmBERT, a Mistral LLM judge, weighted score fusion - and finished in the top 15 out of 80 teams.
-
CEDD Hackathon - Watching the Trajectory, Not the Message: Real-Time Emotional Drift Detection
The second system our team built at Mila's AI safety hackathon (March 2026): CEDD, a real-time monitor that tracks the trajectory of a youth's conversation with a support chatbot - 67 trajectory features, GradientBoosting behind 7 safety gates, adaptive LLM modulation - 90% CV accuracy and 36/36 adversarial tests passing.