Seattle, WA — AI/ML Engineer
Principal, Intelligence Engineering at Slalom. Ten years spanning systems engineering, machine learning, and production agentic platforms — building MCP servers, agentic workflows, and serverless/k8s full-stack applications on AWS, GCP, and Azure.
Case study · 2025 — presentSole developer of a serverless platform for a citizen-science astronomy organization — cut scientific data review from months to days for downstream consumers like NASA and ESA.
Read the case studyEngineering lead for a zero-to-one agentic system performing deep research and report writing. MCP servers plus agentic data-visualization and mapping capabilities.
Core contributor to Slalom’s cloud-agnostic agentic workflow engine — built MCP server capabilities and agent isolation functionality used across client engagements.
Matured a beta GenAI application into a production MLOps platform while halving inference spend; separately built an event-driven personalization service that lifted email clickthrough 85%.
I'm passionate about building AI, ML, and automation-forward solutions to salient business problems. I'm a big-picture thinker who connects AI/ML to tangible business impact, then owns the whole stack to get it shipped — architecture, infrastructure, application code. I develop with Claude Code daily and thrive at the intersection of people, process, and automation.
Outside of work, I love to spend time outdoors! I spent six years as a mountain rescue technician in southern Arizona — 130+ missions with the Southern Arizona Rescue Association. These days I enjoy running (roads and trail) and exploring the Washington mountains. I also enjoy building software in my free time, tinkering with home automation, and gardening.
Engineering lead for an agentic deep-research platform for an AI infrastructure client — hands-on from architecture through production code to delivery.
Core contributor to the Agentic Workflow Accelerator; cut a Fortune 500 client’s LLM inference spend by 50%; led a team of three on an AI-accelerated legacy modernization with Codex. AI SME shaping agentic solution patterns in Azure AI Foundry for enterprise clients.
AI/ML Center of Excellence. Led a team of three running an enterprise RAG system with thousands of users; ~2x improvement in radio signal collision separation with neural networks.
Semantic segmentation for document layout analysis; object detection on synthetic aperture radar imagery; Python tooling for automated reporting.
Automated ~50% of root-cause analysis ticket workflows with AI (~$1M/yr saved); SME for data servers and SQL operations across 30+ sites.
Machine learning and advanced machine learning, computational statistics, statistical models and regression, algorithms for data science, and stochastic differential equations.
Minor in Mathematics. Dean's List with Distinction; senior design team leader for an autonomous-vehicle testbed for GNC algorithm testing.
U.S. Patent 12,615,335 B2 · issued Apr 2026 · Carlson
Real-time speaker recognition for emergency call centers: incoming caller audio is embedded into a latent vector space by a neural network, and a similarity query alerts the dispatcher — while the call is still live — when the voice print matches known hoax callers.
U.S. Patent application 18/439,904 · allowed Jul 2026 · Carlson et al.
A training pipeline that synthesizes colliding AM radio transmissions in the complex I/Q plane — modulation, tuning offsets, path loss, noise — and trains a neural separation model with SI-SNR loss to recover the individual transmissions from the collision.
Proc. SPIE 12529 · Carlson, Liu, Leal, Palermo · doi:10.1117/12.2662948
No datasets exist for AM radio separation, so we generated one: LibriMix speech modulated through a synthetic RF channel. A neural network operating on I/Q data achieves a 98.9% higher SI-SNR than the audio-only baseline — nearly 2× cleaner separation at every tuning offset.
Proc. SPIE 12544 · Beatty & Carlson · doi:10.1117/12.2662974
A convolutional network that detects least-significant-bit steganography in images at up to 96% accuracy — even when the hidden payload is AES-encrypted. Class-activation-map and entropy studies show the network keys on statistical fingerprints the payload leaves in the LSB plane.
Open to interesting problems in agentic systems, applied ML, and the occasional astronomy side quest. The fastest way to reach me: