I build AI-driven products using AI-driven tools — writing production systems at a pace that used to require full engineering teams. Thirty years of architecture and delivery judgment means I know what "production-ready" actually looks like, and I can take your idea from napkin sketch to a well-deployed, operated system fast enough to iterate with you in real time.
That speed doesn't come at the cost of rigor. The same AI-first approach that rapid-prototypes a consumer product also builds the multi-framework compliance platforms, auditable governance pipelines, and HIPAA-grade architectures that regulated industries demand. The cases below cover both ends of that spectrum — and everything in between.
Governance, auditability, and multi-framework compliance aren't constraints on AI delivery — they're design inputs. These projects were built for healthcare, financial services, and technology organizations where every AI output needs a defensible evidence chain and a human approval gate. Production-grade architecture, not demos.
Multi-agent pipeline that analyzes an enterprise application portfolio and produces auditable cloud migration recommendations. Deployed for a healthcare imaging organization: 36 clinical apps, 408 VMs, $24 total AI compute. Every recommendation traceable to its source data by design.
Platform that scans codebases for AI governance gaps across 10 regulatory frameworks, generates compliance artifacts (risk assessments, bias audits, model cards), and enforces gate-based human approval workflows at 5 points in the SDLC. For organizations deploying AI in regulated environments.
An agent that turns a company name into a complete technology diligence brief in under 10 minutes — live news, website analysis, stock data, stakeholder mapping, tech stack assessment, and opportunity framing. Replaces 4+ hours of manual research per account.
Collects 30 days of live datacenter telemetry via PowerShell/Bash agents, builds an in-memory dependency graph, and runs 5 AI simulations: wave rehearsal, blast radius analysis, co-dependency discovery, right-sizing, and TCO projection. Across 1,053 servers, 4 DCs, 2,715 applications.
A complete HIPAA-compliant healthcare AI platform generated from a written spec using Kiro — 9 CloudFormation modules, Bedrock multi-agent swarm, 2 MCP servers on ECS Fargate, Step Functions ETL with PHI masking. Zero IaC written by hand. Demonstrates AI-assisted infrastructure delivery at scale.
The Azure edition of the same spec-driven deployment: 10 Bicep modules, 4-subnet HIPAA VNet, Durable Functions ETL, Cosmos DB multi-API, Azure AI Search vector indices, Container Apps MCP servers, and a GPT-4o multi-agent swarm. Includes an offline demo mode deployable in 2 minutes with no Azure account.
The enterprise AI governance platform deployed to AWS with zero forked business logic. A cloud adapter layer maps Cosmos DB → DynamoDB GSI, Key Vault → Secrets Manager, Azure AI → Bedrock — all invisible to the application handlers. Same features, different cloud, no duplicated code.
The original enterprise AI governance platform, productionized. Kiro generated the full IaC (4 Bicep modules), CI/CD pipelines, a SHA-256 blast radius audit tool, and a unified seed runner that replaced 18 numbered setup scripts. Zero business logic was changed — only the deployment scaffolding was built around it.
Park Whisperer started as a training exercise — a domain I knew well (Walt Disney World) used to practice building AI systems. Over time it evolved into something else entirely: a real, working product with users, live data pipelines, an ML stack, automated content operations, and a conversational agent. It demonstrates what's possible when you take a business idea from zero through a complete AI ecosystem — and operate it reliably in production for under $100 a month across three clouds.
Full platform architecture →Production GCP pipeline that fuses METAR surface observations, NWS forecasts, and upper-air radiosonde soundings in BigQuery. ML scoring maps current conditions to per-attraction OPERATE / MONITOR / CLOSE recommendations and publishes a live JSON API to GCS every 15 minutes, consumed by the agent and dashboards.
Kiro transformed a working TypeScript governance library (136 tests, 6 components) into a full multi-tenant SaaS service — API key auth, PostgreSQL + Redis tenant isolation, Git-webhook policy hot-reload with zero-downtime blue-green index swapping, and a hybrid SDK with automatic agent/library fallback. 13 tasks, no downtime.
ETL pipeline and SLM-routed query engine over 1.6M+ theme park operational records. A fine-tuned Phi-3 Mini classifier routes queries to one of 3 strategies (vector, numeric, hybrid). Deterministic fast paths bypass AI inference entirely for wait-time and outage questions — no latency, no hallucination.
Scheduled pipeline that generates and publishes theme park social content with no human involvement. Intent classification → parallel knowledge pre-fetch → LangChain agentic loop with 30+ tools → 3-pass generation → publish. 10 content types, zero hallucinated facts, grounded entirely in live operational data.
Four AI models with distinct roles collaborate on each content piece. Phi-3 Mini classifies intent. Haiku runs the agentic tool loop (fast, cheap). Sonnet writes the final output — replacing Haiku after model-specific testing revealed it lacked domain knowledge for the subject matter. Each model has a persona and guardrails.
Five systems take a Claude-written script from text to a live social post with zero manual steps. ElevenLabs per-character timestamps drive frame-accurate word-pop captions. Pillow composes brand overlays. FFmpeg encodes. An API publisher posts to three platforms. The only human input is approving the script.
Conversational agent that answers questions about park operations in plain language. Uses the same Azure Function, SLM classifier, and 15+ live data tools as the content pipelines — the agent and content factory share infrastructure. Designed for responses short enough to paste directly into social DMs and comments.
Containerized ETL job running every 10 minutes that collects live attraction status, wait times, Lightning Lane sellout events, and schedules for 286 WDW entities and writes them to PostgreSQL. Feeds every downstream consumer: the RAG, the agent, and all content pipelines. The foundation of the platform.
How the platform's analytics store went PostgreSQL → Cosmos DB → back to PostgreSQL. A Cosmos DB monolith developed a critical hot-partition problem at 5.9M documents. The fix cost $394/month with no benefit. This documents the decision process, the migration, and the six architectural lessons — including when not to use a document store.
8 years of Central Florida METAR surface obs fused with twice-daily radiosonde soundings (CAPE, Lifted Index, K-Index) trains 6 scikit-learn models that nowcast thunderstorm, precipitation, fog, and venue-impact risk every 20 minutes. A spatial boost via BigQuery ST_DISTANCE amplifies probabilities when confirmed active storms are within 15 miles.
Ingests GOES-18 Geostationary Lightning Mapper data every 20 seconds, applies a two-stage spatial filter to attribute flashes to specific parks, and drives a 5-state storm machine (CLEAR → SEVERE) that fuses satellite data, 17-station METAR surface obs, NWS alerts, and SPC categorical outlook to compute per-ride closure risk.
Each of these preceded a production component. They're documented because the path matters — they show what failure modes look like, what tradeoffs led to the next iteration, and how the production system got to where it is.
Global LightGBM model across all WDW attractions using a recursive multi-step loop — predicted values fed back as lag features for subsequent horizons. Seven forecast intervals from 15 minutes to 3 hours. One of 8+ iterations in the wait-time forecasting R&D track, each informing the next architectural decision.
Per-attraction time-series forecasting with one Prophet model per ride, 50+ external regressors, and a 24-hour horizon written back to BigQuery. Reached v8 before being shelved — rolling window features unavailable at inference time degraded accuracy significantly.
Three-cloud exploration of Bedrock Agents and custom action group patterns. Wired three weather tools to a GCP Cloud Run backend via two integration architectures: direct Lambda action groups and Azure Logic Apps as cross-cloud backends. The tool design patterns and OpenAPI schema approach carried forward directly into the production agent.
Earlier agent architecture using the Azure OpenAI Assistants API (threads/runs model). Persistent assistant object, 16 function tools, server-managed thread state, Cosmos DB vector search over 9 knowledge types. Directly preceded the production Park Agent Chat.
"Walt" — a central host agent — coordinated seven specialized sub-agents, each named for a Disney character and owning a park knowledge domain. Two custom FastAPI microservices ran alongside n8n. Iterated through 7 Docker Compose versions before being replaced by the Azure Functions architecture.