Federal Reserve Bank of New York
Anti-money-laundering entity resolution and due diligence, screening the Bank against reputational risk, with Neo4j graph analytics resolving entities and relationships across counterparties.
AI ArchitectActive DoD Secret Clearance
MatthewR. Pollard
Fourteen years building production machine learning. The interesting part is knowing when it's wrong.
- 14
- Years of production ML
- 7
- Engineers founded and led
- 20%
- More side effects detected at J&J
- 99.9%
- Pipeline uptime at AI C-Suite
01Record
Where being wrong is expensive.
Production machine learning for institutions that are regulated, audited, and unforgiving of error.
Walmart and eBay
AML detection systems that carried both firms through regulatory approval, including Walmart's classification as a Money Services Business. Technical lead on Walmart's; engineer on eBay's.
Johnson & Johnson
Drug-safety signal detection over reported adverse-event data, increasing side-effect detection by 20%.
Department of Defense
Cleared work on the F-35 program at MCAS Cherry Point and in AWS GovCloud with CUI workloads: model evaluation, agent systems, and inference on hardware with no network connection.
02Now
What I build today.
Fine-tuning, evaluation, agents, and inference close to the metal.
Model evaluation
Benchmark harnesses, regression suites, and release quality gates. Established the practice where none existed.
Multi-agent systems
Agent orchestration and autonomous scheduled agents inside AWS GovCloud, with operator interfaces for directing and auditing runs.
On-device inference
Gemma 4 on LiteRT-LM and Ollama for laptop, browser, and mobile targets, including hardware with no network connection.
Retrieval and search
Retrieval-augmented generation over FAISS vector indexes fed by ONNX Runtime embedding pipelines, knowledge-graph retrieval, and cross-encoder reranking.
Fine-tuning
Pipelines for Mixture-of-Experts models (Gemma 4) and NVIDIA Nemotron: dataset curation, synthetic data generation, training orchestration.
GPU kernels, CUDA and Rust
Fused kernels (Liger Kernel, FlashAttention), GPU environments tuned at the CUDA driver and runtime level, and latency-sensitive async and parallel Rust services.
03Specification
Technical specification.
- Languages
- Python, Rust, C# / .NET, SQL, Java, TypeScript, SAS
- ML & Deep Learning
- PyTorch, fine-tuning (Mixture-of-Experts, LoRA/PEFT), dataset curation, synthetic data generation, training orchestration
- LLM Systems
- RAG, embeddings, FAISS, cross-encoder reranking, sentence-transformers, knowledge graphs, NLP, LangChain, Hugging Face, OpenAI APIs, NVIDIA Nemotron, Gemma 4
- Agentic Systems
- Multi-agent orchestration, autonomous scheduled agents, MCP (Model Context Protocol), tool use, operator interfaces
- Evaluation
- Benchmark harnesses, regression suites, release quality gates
- MLOps & Experimentation
- Docker, Kubernetes (via Google Anthos), CI/CD, custom-built experiment tracking and pipeline orchestration, Weights & Biases, SageMaker experiment tracking
- On-Device & Edge Inference
- LiteRT (formerly TensorFlow Lite), Ollama, ONNX Runtime, quantization, WASM and mobile deployment
- GPU & Performance
- Fused kernels (Liger Kernel, FlashAttention), NVIDIA CUDA (driver and runtime), async and parallel systems
- Web & APIs
- React, Next.js, Svelte, SvelteKit, FastAPI, Fastify, REST
- Cloud
- AWS and AWS GovCloud (SageMaker, S3, EC2, ECR, Lambda, Glue), Google Cloud (Compute Engine, GPU training VMs, Anthos), Azure
- Databases & Data Engineering
- PostgreSQL, Oracle, Neo4j (graph analytics for financial fraud), Spark (via AWS Glue), ETL, SAS Viya
04Service record
Fourteen years, no gaps.
- Jan 2026 – Aug 2026Vergence, LLC
Senior AI Engineer
- Established the model evaluation practice: benchmark harnesses, regression suites, and release quality gates.
- Designed multi-agent orchestration and autonomous scheduled agents inside AWS GovCloud, with operator interfaces for directing and auditing runs.
- Built on-device Gemma 4 inference on LiteRT-LM and Ollama, and fine-tuning pipelines for Mixture-of-Experts models and NVIDIA Nemotron.
- Nov 2024 – Nov 2025SteerBridge Strategies
F-35 AI Engineer
- Delivered analytics and ML tooling to 2nd Marine Aircraft Wing officers at MCAS Cherry Point on the F-35 program.
- Trained PyTorch models for maintenance scheduling, parts availability, supply-chain forecasting, and resource allocation.
- Stood up the supporting AWS GovCloud pipelines in Glue, with model deployment and performance monitoring.
- Jun 2023 – Nov 2024AI C-Suite
AI Manager and Lead Engineer
- Led five engineers building AI and automation systems for enterprise clients; owned architecture and delivery.
- Ran PyTorch pipelines on Google Cloud at 99.9% uptime with automated continuous model updates.
- Created DepoLite, an agent-based platform that turns deposition transcripts into litigation-ready summaries, and the Norman Schall & Associates court reporting operations platform.
- Jun 2019 – Jun 2023ZETI AI
Founder and Principal AI Engineer
- Assembled and led a seven-engineer team training large-scale LLM and diffusion models on NVIDIA A100 clusters.
- Shipped Zeti Dreams to the App Store and Google Play, and a web DreamBooth platform for user-trained Stable Diffusion models.
- Published reinforcement learning work: Destino, Zarena, AI Library, and contributions to muzero-general.
- Jun 2012 – Jun 2019SAS Institute
Senior Technical Architect
- Architected ML systems for the Federal Reserve Bank of New York, Walmart, eBay, Johnson & Johnson, EMC, and State Farm.
- Built the transaction monitoring pipelines in SAS Viya and Oracle across billions of rows of retail financial data.
- Engineered the drug-safety signal detection system at Johnson & Johnson and the AML entity resolution system at the Federal Reserve Bank of New York.
05Built and shipped
Open source and products.
Open source
Destino
Complete MuZero with training and an API, driving a Rust Monte Carlo tree search to play any turn-based game in OpenAI Gym environments.
Zarena
Reinforcement learning environments for OpenAI Gym written in Rust and compiled to WebAssembly.
AI Library
Foundational transformer, diffusion, and reinforcement learning notebooks across frameworks, including Monte Carlo and MuZero.
muzero-general
Contributor to the reference open-source MuZero implementation that Destino forks.
Products
Zeti Dreams
Consumer generative-image app trained on A100 clusters and shipped to iOS and Android.
Norman Schall & Associates
Court reporting operations platform for a Los Angeles firm, with DepoLite turning deposition transcripts into litigation-ready summaries.
06Credentials
Cleared, certified, degreed.
Secret
Active DoD clearance
Held continuously through the SteerBridge and Vergence roles. CUI workloads in AWS GovCloud.
CompTIA Security+
Issued 2026. DoD 8140 IAT Level II.
University of North Carolina Wilmington
B.S. Computer Science. Thesis: Predicting the US Dollar Index with Recurrent Neural Networks.
English and Spanish
English native. Spanish B1.
Let's talk.
The fastest way to reach me is LinkedIn.