I design and build AI automation platforms end-to-end: voice agents, LLM orchestration pipelines, enterprise web apps, and real-time event-driven infrastructure. This is the work.
Not demos. Not side projects. Systems handling real traffic, real users, and real consequences every day.
FastAPI web application with embedded Claude AI agent giving 200+ non-technical staff one-click access to 64 automation tools. Full role-based access control, audit logging, and Docker deployment. Replaced manual CLI workflows company-wide and enabled the business to scale operations without adding headcount.
Real-time AI voice state machine controlling behavior for 600+ shop locations via n8n and Genesys Cloud APIs. Handles 50,000+ monthly conversations across inbound, outbound, and overflow flows. Replaced static per-location IVR configuration that previously required manual updates for every change.
Fully productized AI voice agent handling appointment booking, availability checks, call memory, and human escalation. Built on a 6-webhook event pipeline connecting Vapi.ai, n8n, Cal.com, and Google Sheets. Deployed for client businesses as a turnkey AI front desk.
SIP/PCAP packet analysis combined with Genesys CDR API data to reconstruct complete call paths, identify failure points, and surface audio quality issues. Reduced call investigation time by 87%. Built a 2,071-line IVR parser that enumerates every distinct caller path through any flow definition.
Automated LLM evaluation infrastructure using parallel LiveKit call agents orchestrated by Claude. Auto-generates test cases directly from IVR flow definitions and runs 20 concurrent synthetic calls to validate behavior before any change reaches production. Previously, IVR changes were tested manually in live environments.
Directive/Orchestration/Execution pattern that separates LLM decision-making from deterministic execution. The core insight: 90% accuracy per step compounds to 59% success over 5 steps. This pattern keeps AI in the decision layer only. Self-annealing loop: every runtime failure fixes the script and updates the directive. The system gets stronger on each failure.
Multi-model orchestration system ingesting FRED macroeconomic data, Finnhub financials, GDELT news sentiment, and ACLED conflict data. Claude Opus orchestrates deep signal analysis; Claude Haiku handles batch preprocessing at scale. Built to demonstrate multi-source fusion and intelligent model routing based on task complexity.
Spanning contact center operations, voice AI, enterprise provisioning, analytics, and sales automation. A representative sample below.
From internal enterprise tools to client-facing business websites. All designed, built, and deployed end-to-end.
Multi-user FastAPI web app with Claude AI agent. 64 tools available to 200+ non-technical staff. Role-based access, audit logging, Docker deployment on VPS.
Business automation services site with Cal.com booking integration, mobile-responsive design, and live testimonials. Built and deployed from scratch.
Client business website for a catering company. Custom design, mobile-first, inquiry form integration.
Client website for a cleaning services business. Service area pages, booking flow, brand-matched design.
Client website for a contracting business. Project gallery, service descriptions, and lead capture form.
Custom webhook router deployed on Modal. Maps inbound requests to directives and executes Claude-orchestrated workflows. Powers the automation layer across multiple production systems.
I take on a small number of consulting engagements alongside my full-time work. If you're building something that needs serious AI infrastructure, let's talk.