The Integration Pattern
20 years building systems that bridge different contexts and abstraction levels
When I read Google's Nested Learning paper, I didn't see a novel optimization algorithm. I recognized an architecture I'd been building for two decades in other forms.
The same pattern appears everywhere I've worked: language instruction in Paris, business English at Rice, beekeeping IoT, cross-platform gaming APIs, healthcare ML. Different domains, same underlying challenge—how do you let different systems communicate without losing their distinct characteristics?
Implementing frontier research in one day isn't about coding speed. It's about recognizing integration patterns I've built before in other contexts.
Why I Can Move Fast
Most ML engineers either understand research OR ship production systems. I do both because I've been building integration layers between different systems my whole career.
Reading papers and implementing in PyTorch uses the same skill as building API bridges between Minecraft, Roblox, and Fortnite. Both require understanding how different systems represent information, then building translation layers that preserve what matters.
When I hit the normalization strength issue in my collaborative nested learning implementation, I recognized it immediately—it's the same problem as the gaming APIs. How do you let Minecraft, Roblox, and Fortnite communicate without making them all behave the same? How do you let fast, medium, and slow optimizers share knowledge without losing their distinct characteristics?
This is pattern recognition developed across domains, not just coding speed.
The Domains
Paris (2005-2009)
Founded JLS Langues, a language instruction business coaching French banking executives at BNP Paribas, Société Générale, and other major institutions on American business communication. Not just language—cultural code-switching between French and American business contexts. Teaching at different abstraction levels: pronunciation, grammar, discourse, cultural context. Each level requiring translation between complexity systems.
Rice University ESL
Taught business English to extraordinarily diverse professional learners— Taiwanese and Chinese students in the same classroom, Saudi women and men discussing American culture, Japanese and Korean learners collaborating, Turkish and Russian students together. Managing cultural dynamics while teaching language structure. Navigating complexity across multiple dimensions simultaneously.
Beekeeping (2021-present)
Founded Jay's Bees with Mellifera voice-first hive management app. Built bidirectional voice interface (Web Speech API + AWS Polly TTS) for hands-free field operation—beekeepers can't use touchscreens in thick gloves. GPT-powered NLU extracts structured data from natural speech. The inspection/treatment/monitoring workflow mirrors clinical healthcare, giving me intuition for healthcare data challenges.
Cross-Platform Gaming APIs (2024)
Built integration bridges between Minecraft, Roblox, and Fortnite to teach my 9-year-old about APIs. Connected weather APIs and LLM APIs to create cross-platform quests communicated by NPC guides in each universe. Same architectural challenge as ML: how do you let different systems communicate without losing their distinct characteristics? This is where I first solved the normalization problem that appeared again in nested learning.
Music
Studied saxophone under Jeff Coffin (Dave Matthews Band, Béla Fleck), working in the tradition of Sonny Rollins' theme-based improvisation— developing melodic ideas through variation rather than Coltrane's more syntactical approach. Real-time pattern recognition and variation.
The thread: Every domain involves building bridges between systems that operate at different levels or in different contexts. The specific technology changes; the integration pattern doesn't.
Current Focus at HCA Healthcare
Production ML systems serving 180+ hospitals with 44M+ patient encounters annually. Systems where reliability matters, not just research demos.
- Kaizen AI — Innovation management system that captures clinical and operational improvement ideas, using ML and GenAI to enhance, validate, and prioritize them. Executive visibility into innovation pipeline across the organization.
- Clinical RAG Platform — Decision support across 180+ hospitals using retrieval-augmented generation. Production system with real clinical impact.
- Continual Learning Research — Implementing and extending Google's NeurIPS 2025 Nested Learning for production healthcare use. Patient safety requirements make this more than academic.
- ML Infrastructure — Platformizing AI deployment with IaC patterns: prompt registries, post-prediction data gathering, eval pipelines, and versioning. Making continuous learning a first-class concern.
- Conductor — Language-agnostic automation orchestration in C#. Handles provisioning and coordination across multiple RPA and scripting environments. Deployed on GKE with CloudSQL.
In healthcare, forgetting is dangerous. A clinical AI that forgets baseline knowledge when learning new protocols isn't just academically interesting—it's a patient safety issue. I build systems where learning is safe.
Background
Transitioned from traditional software engineering to ML engineering. Completed comprehensive ML education (Andrew Ng's specialization, self-study). Built production systems as technical demonstrations (Mellifera voice-first app, agentic platforms). Developed deep intuition for healthcare data challenges through beekeeping's inspection/treatment/monitoring workflows.