Checkpoint path
Memory and local or peer NVMe handled the fast path. Blob provided durable recovery without making every checkpoint wait on object storage.
Systems, architecture, and code
Here are the systems and the mechanism behind each one.
The work ranges from a three-person school website to storage paths for large AI workloads. I have separated what I designed, what I built, what a larger team delivered, and why the architecture mattered.
2017–2025 · Microsoft Azure Storage
The GPUs were expensive. Storage could not be what left them idle.
From 2017 through 2025, I led Azure-side storage work supporting OpenAI and Microsoft AI. The architecture had to keep training data and checkpoints moving, put capacity near the compute, and recover cleanly when something failed.
Namespace and routing
Scaled Accounts let one namespace span multiple storage clusters. Placement, routing, and rebalancing had to keep working as capacity was added.
Data placement
Central repositories held training data and checkpoints. Regional storage near the GPUs was pre-staged before compute turn-up, with service-side movement over Azure’s private WAN.
Operating loop
Capacity reviews, SLO/SLA telemetry, regression detection, Sev0/Sev1 response, RCAs, and roadmap changes kept the architecture connected to production.
I also led Azure Blob’s storage and data workstream for Project Fairwater across compute, networking, hardware, capacity, supply chain, service operations, performance, and reliability.
These were team outcomes. My role was architecture, product, capacity, and operations underneath the models, not model research.
01
2025–2026
Inference data architecture at VAST Data
KV cache is capacity, write load, latency, and hardware life.
I worked on the data architecture and economics behind AI training and inference.
The work connects concurrency, context length, throughput, utilization, and write patterns to GPU memory, host memory, flash, network, persistent storage, SSD endurance, and total cost.
02
2024–present
Personal agent fleet
A private developer platform that has to keep working.
I run a pod-per-repository fleet of coding agents on a live Kubernetes cluster.
GitOps defines the desired state. Each repository gets an isolated runtime and worktree, while model routing, observability, and explicit controls make failures visible and recovery repeatable.
Sources
03
2024–present
Verdify
AI can propose a change. Firmware still controls the greenhouse.
I am building an AI-assisted greenhouse control loop with my adult son, James.
The ESP32 state machine owns safety and actuation. AI can tune only bounded parameters, and telemetry plus a decision ledger tells us whether each change worked.
Sources
04
2017–2025
Azure AI storage architecture
Training data and checkpoints had to move without leaving the GPUs idle.
I led Azure-side storage architecture, product, capacity, and operations work supporting OpenAI and Microsoft AI.
The system combined a memory and NVMe checkpoint path with durable Blob recovery, one namespace across storage clusters, placement and routing near compute, pre-staging, service-side data movement, and an operating loop tied to production.
05
2015
Azure high-IOPS quickstart
A small public artifact that anyone can inspect.
I contributed a 32-disk VM and storage-pool Azure Resource Manager template.
The original pull request added five files and 614 lines, and the README follow-up also merged. Azure revised the code later, while the successor metadata still credits jvallery.
Sources
06
2007–2016
MineMyMail
I built the utility I needed, then turned it into a product.
MineMyMail extracted, deduplicated, enriched, and exported contacts from IMAP mailboxes.
It began as a Gmail address-extraction tool and grew into a paid product. It was small and imperfect, but people were still using it in 2016.
Sources
07
2006–2008
NewsPicker → Mine Feed
A working social-news and RSS experiment.
I launched and operated a Pligg-based site with accounts, submissions, voting, queues, and RSS.
I customized and ran the Pligg-based system rather than writing the underlying engine from scratch. NewsPicker later became Mine Feed.
08
2001–2008
Brio Solutions
My first company ran beside my day job.
I built and hosted LAMP-stack sites and web applications for small organizations.
I handled the code, hosting, deployment, and customer work. It was where I learned how quickly an architecture becomes an operating problem once another person depends on it.
Sources
09
1996–1998
The first web chapter
Publishing, astronomy, Perl, and Linux before I finished high school.
The surviving record includes a student newspaper byline, a three-person ThinkQuest site, an Internet Guide page, and public troubleshooting threads.
Those projects taught me to build something real, inspect where it failed, ask specific questions, and change the system until it worked.
Current projects
I still test architecture by operating it.
I run the agent platform and build Verdify with my adult son. Both systems have real state, failure modes, and a cost when the design is wrong.