Donald Galliano III

Sleep technologists pass their national board at 38%. Training is bottlenecked by lab access and random patient volume. SleepSimSocial is a flight simulator that fixes that.

• Built a synthetic patient engine where comorbidities drive physiology—COPD patients present with complex central apneas, cardiac patients exhibit arrhythmias mid-study—and clinical decisions dynamically alter patient state in real time
• Natural language scenario generation: instructors describe scenarios in plain English ("generate a narcolepsy MSLT with cardiac arrest at hour 3") and the system builds, scores, and grades automatically
• Partnered with international workforce solutions company (1,000+ employees, 60+ locations) and deployed pilot at a CAAHEP-accredited program with 11 students
• Patent counsel identified novel patentable concepts; inbound interest from children's hospital for pediatric simulation expansion

A 5-tier cryptographic reasoning competition designed to systematically eliminate classes of AI approach, from prompt engineering to combined multi-layer attacks.

• Each tier targets a different capability: prompt engineering, traditional cipher analysis, fixed-base mathematical reasoning, sequential computation, and combined multi-layer attacks
• Parametric puzzle generators with tunable difficulty scaling across substitution ciphers, self-modifying equations, variable-base arithmetic, and recursive tree-based encryption

Patient deterioration prediction system classifying ED vital signs into Normal/Warning/Crisis/Death states, where clinical domain expertise produced the largest single-class accuracy improvement of any architecture change.

• Engineered a 38-feature matrix including pharmacovigilance drug interaction profiles and latent consciousness inference; 6 hand-coded sigmoid thresholds from RPSGT expertise formed a clinical ECG gate that outperformed every architecture change
• Extracted 42 ECG morphology features per encounter—30 statistically significant between Warning and Crisis (p<0.05). ST elevation 29x higher in Crisis. Discovered Warning and Crisis depend on fundamentally different physiological mechanisms
• Three-model ensemble (GP + BNN + LSTM) achieved 88.26% accuracy. LSTM broke through the Crisis ceiling to 81.21%, reducing early-encounter Warning miss rate from 76.9% to 10.1%

An automated red team layer that breaks features before and after they ship, orchestrating multi-agent workflows across browser and terminal environments.

• Chrome extension and WebSocket bridge orchestrating multi-agent workflows across Claude.ai tabs and Claude Code terminals
• Red team adversarial agentic loop that automatically discovers vulnerabilities and edge cases, functioning as an automated penetration testing layer across all production projects

LeetCode meets ML, but fun. A programming-puzzle sandbox disguised as a trading card game where players solve problems by building card logic from composable computational keywords.

• 8 composable keywords map to computational paradigms; players construct card logic to solve challenges that scale in complexity
• ML ensemble (K-means, XGBoost, prototypical networks, BNNs) detects degenerate strategies and patches balance in real time—no manual tuning needed
• Rust core with PyO3 bridge to Python ML stack for performance-critical game logic with ML-driven balance on top