Recent projects, and a few I made for fun.
Client systems sit next to personal tools. Each write-up is the architecture, the ugly constraints, and the part that was actually interesting to solve — not a screenshot gallery.
Azure Local Operations Platform
Not a normal monitoring dashboard: an evidence-first platform for 152 Azure Local clusters across five walled-off environments. PowerShell collectors, WMI fault listeners, Windows Event Forwarding, and Azure APIs feed SQL Server 2025 as the processing engine, behind typed APIs and a desktop-grade React dashboard.
The Forge
Governed PowerShell automation with no shell exposed: operators launch allowlisted, hash-pinned commands after an explicit blast-radius check; per-environment Windows runners claim work from SQL, stream output to a live terminal, and leave a log, result JSON, and ZIP for every executed run.
Ask ALDHI
An AI assistant that investigates with evidence: 38 typed tools over 15 catalogued sources, a sandboxed Code Mode, and structured answers that cite what they used and label what's report-backed. No model-written SQL, no secrets in context. Passed 30/30 live eval prompts in June.
Event Intelligence
A dictionary of 114 researched event providers and frozen, replayable analysis runs over ~17M forwarded Windows events. It ranks what deserves attention — and refuses to treat silence as health.
AlphaForge — risk-first trading platform & AI Lab
A risk-first trading platform for crypto and ETFs. Strategies and an AI Lab only propose; a fail-closed risk engine decides; every fill lands in an append-only ledger. Paper trading end to end on Azure; live trading stays off. Code stays closed.
design-doctor
A policy-driven design-quality scanner for React 19, shadcn/ui, and Tailwind v4. Sixteen rules, a design policy the team writes down, SARIF and GitHub output, and an agent skill for coding assistants. Pre-release MVP; Tailwind language-service analysis and motion-token drift are next.
Azure Local Redeployment Console
A scrappy PowerShell console I built in about a day, then hardened over the program: the runbook in order, per-site progress, dry runs, manual checkpoints, and commands logged by site. That's how I cleared 96 Azure Local redeployments in two weeks, solo.
Fintech Analytics & Automation
Invoice PDFs in, reconciled numbers out. Email-fed Python Azure Functions run custom-trained Document Intelligence models, about 90 T-SQL procedures link invoices and match deposits and ACH payouts, and a Next.js 15.5 dashboard turns the result into executive, cash, and risk views.
Fiber Optic Project Management
Operations platform that moved a fiber engineering firm off spreadsheets: KMZ/KML route maps, segment workflows, POs, three-reviewer invoicing, payments, and payroll on Next.js 16 and SQL Server. Serializable payments, database triggers, allowlist-only Microsoft sign-in, and Admin-only financials.
Video Frame Analysis Pipeline
Python GPU pipeline and terminal app that scores every video frame with smartphone-trained quality models, keeps every score in SQLite, and exports the best stills through a hardened FFmpeg path. Rebuilt in August 2026 after a benchmark caught the old models favoring blur.
Happy to walk any of these end-to-end.
The case studies cover the architecture and engineering decisions; a live tour adds the operational context that cannot be published.