Content Integrity for all of Meta's surfaces (Facebook, Instagram, Threads, Ads, Commerce, WhatsApp Business)

Multimodal e2e AI models for addressing "high prevalence problems", collectively reaching a userbase of 3.58 billion daily active people (DAU).

Optimizing the innovative technology that powers the world's first Generative Al products, with a special focus on Al safety and responsible deployment for our billions of consumers.

Improving the way AI Agents create performant C++ binaries.

Developing and launching tooling for creating embodied, environment-aware, AI-powered Non-player characters (NPCs) for the Horizon Worlds platform, enabling creators to build LLM-powered characters into virtual worlds.

Building flagship games to demonstrate the aforementioned capabilities with real time AI color commentators, AI in-game shopkeepers, AI in-game sidekicks, and AI bosses for social deduction games (featured at Meta Connect 2024 & 2025), available to millions of MAUs.

Training and deployment of LLaMa-based LLMs and Text-to-Speech (TTS) models for low latency (sub-300ms) real-time interactions, as well as zero-shot text-to-TTS-voice generation, Vision-Language Models (VLMs), Action-Language Models (ALMs), and other multimodal language models (MLMs)

AI NPCs overall generated a 2% incremental increase in user time spent, accounting for over 300,000 hours annually.

Improving RAG for memory and conversations, allowing for coherent longer term real-time conversations with users while saving 85% on LLM memory.

Safety and integrity systems for ensuring compliance with state-level, federal, and international laws.

Managed ML research and deployment strategies for 50+ offerings in B2B and B2C.

Built 100+ ML pipeline analytic processes to decrease clients’ costs by 20-45% (with extreme examples of cost savings being as high as 2,500x).

Worked with customers in Music, Law Envirocement, Biotech, VR/AR, including a TechCrunch Disrupt Product Launch that generated 10k users.

Created templates for resolving CUDA and NVIDIA driver combination environments in AWS EKS GPU nodes, in order to reduce user start-up time compared to services like Google Colab for applications like protein-folding, inverse-folding, diffusion-based molecular docking, image segmentation, and high-dimensional visualization.

Architected bioinformatics-specific code-completion-assistant with emphasis on LLM-agnosticism, user data security, user customization, cost-effectiveness, and verification of AI output on indexed real-world data by using AWS Sagemaker, reduced weight precision inference, optimized vector search, and ML security principles outlined in my recently-published O'Reilly book (see publications).

Worked with interdisciplinary team of stakeholders to carry out customer discovery research to identify pressing pain points and productivity bottlenecks among biomedical researchers in both academia and industry (both public and private companies).