About
I'm Steven W. White. This site is named for Key Largo, the first island on Florida's Overseas Highway (A1A) and where many of my favorite drives begin.
Somewhere between the mainland traffic and open water, there is a stretch of road where the noise drops away and the thinking clears. No meetings stacked back to back, and no pressure to ship a slide deck before understanding the problem. Just a question worth testing, a terminal, and time to work from first principles. I write here from that headspace whenever I can find it.
The Journey
I have spent more than 25 years building production software and distributed infrastructure, starting in telecom networks before moving into computer vision, deep learning, and applied AI systems.
In 2009, I founded ImageVision to tackle automated image recognition and content safety for large online platforms. We raised $12.5M, delivered production systems for Google, Facebook, Apple, Samsung, and In-Q-Tel, and scaled pipelines inspecting billions of images. Two Hat Security acquired ImageVision in 2019, and Microsoft acquired Two Hat in 2021. That work continues to help protect communities across Xbox, Minecraft, and MSN.
Following the acquisition, I joined NVIDIA as a Deep Learning Solutions Architect, focusing on TensorRT quantization, low-latency GPU serving, and industrial computer vision at the edge. In 2019, I joined AWS, spending seven years as an AI/ML Solutions Architect and engineering lead designing retrieval architectures, model-serving pipelines, and multi-agent security systems.
Today, my work sits at the intersection of systems architecture and hands-on software engineering: building multi-agent systems, evaluation harnesses, retrieval pipelines, and production ML workloads that hold up under real latency, reliability, and cost constraints.
What Production Teaches You
Across startups, hardware labs, and cloud platforms, the failure modes in applied AI repeat with surprising consistency. A few principles guide how I build and how I write here:
- Demos and production live on different planets. A notebook that works on ten hand-picked prompts tells you very little about tail latency, malformed inputs, auth boundaries, or cost at 1,000 requests per second.
- Measure before you opine. Whenever an architectural choice comes up, run the benchmark and inspect the numbers. A 50-line measurement script beats a week of debate.
- Build the evaluation harness before tuning prompts. If you cannot measure precision, recall, groundedness, and latency regressions automatically on every commit, every change is guesswork.
- Keep contracts typed and explicit. Clear schemas at service and tool boundaries prevent more production incidents than clever orchestration ever will.
- Design for the team that owns it next. The best systems are simple enough that another engineer can debug them at 2 a.m. without calling the original author.
Building FoilFox
Outside my day job, my son and I are co-building FoilFox, a computer vision application for trading card collectors. Point a phone camera at a card in your hand, identify the exact printing and foil finish, check live market pricing, and organize your collection into digital binders.
FoilFox brings me right back to the physical constraints that make computer vision rewarding: glare on holographic foil, sub-second mobile inference budgets, near-duplicate set reprints, and building an interface simple enough that collectors enjoy using it at a trade table.
What You Will Find on largo.dev
Every guide on this site is written the way I wish technical tutorials were written when I was learning a new stack: first-principles intuition, measured outputs from real hardware, and complete code you can clone and run end to end. Right now the library includes 27 hands-on tutorials, 5 technical analyses, and interactive engineering tools across five active tracks:
- Production ML (9 guides): GPU VRAM sizing, MLflow and Langfuse experiment tracking, ML CI/CD gates,
torch.compileand ONNX serving benchmarks, drift detection, security, and time-series failure prediction on 12.2M+ drive-days. - Embeddings & Reranking (5 guides): Training sentence embeddings in PyTorch, scaling bi-encoder search, cross-encoder reranking, multimodal CLIP search, and cross-attention fusion over text and tabular features.
- AI Agents (7 guides): Typed state-graph orchestration (Google ADK, LangGraph, AWS Strands), consolidated long-term memory, and tool-using agents wired to live external APIs.
- Transformers & Reasoning (3 guides): DeepSeek V3.2 and R1 architectures, Multi-head Latent Attention (MLA), sparse Mixture-of-Experts routing, and reinforcement learning for reasoning.
- Retrieval Systems (3 guides): Vector indexing, metadata filtering, hybrid retrieval, and domain-specific semantic search.
- Interactive Engineering Lab: Three browser-based simulators covering GPU VRAM and KV-cache scaling (dense GQA, Hybrid Sliding/Global MoE, and DeepSeek MLA from 2K to 128K context), two-stage BM25 + dense RRF retrieval and cross-encoder reranking latency SLOs, and stateful multi-turn agent context and slot memory consolidation.
- Articles & Constellation (5 essays): Architectural comparisons, annual practitioner reviews, engineering career roadmaps, and an interactive concept map connecting the full library.
Education & Patents
- M.S. in Data Science, The University of Texas at Austin
- B.S. in Computer Science, The University of Texas at Dallas
- 2 U.S. patents in computer vision and machine learning
Connect
You can find my public tutorial repositories on GitHub (@steve-largo), connect on LinkedIn, or send a note directly through the contact page.
largo.dev is an independent personal notebook. Everything published here reflects my own views, benchmarks, and side projects, and does not represent any current or past employer.