Marquetta – Python, OpenAI, AI agent development

A browser-based live production control system for an interactive virtual show. The system connects a host-operated control deck, OBS scenes, audience voting, scoreboard views, QR-based participation, and show-state overlays into one coordinated experience. It was designed so a non-technical host could manage scene changes, guest moments, audience interaction, scoring, and show flow from a simple control surface, while the audience participated from their phones.

• Translated a live-show concept into a structured technical workflow with host, guest, audience, and operator views.
• Designed the show-state model for scenes, rounds, scores, QR voting, and audience participation.
• Coordinated OBS scene design with browser-based overlays and a host-facing control deck.
• Implemented a real-time voting and scoring flow with QR-code entry points and synchronized scoreboard views.
• Defined operator-safe controls so the host could move through show segments without manually managing every technical layer.
• Tested the production views and identified readiness gates before public use, including full-run rehearsal, audience voting test, recording review, and show-flow validation.

A decision-support product that replaces manual opportunity research. It ingests listings from nine sources, classifies each one for location eligibility, scores it against a configurable rubric, states in plain language what to do next and why, and tracks status through to outcome. Explainability is a hard requirement — the product never asks a user to trust a score it cannot itemize.

• Designed the scoring rubric as external configuration rather than code, so scoring can be retuned without a deploy.
• Built ingestion across nine sources including REST APIs, RSS, and two ATS providers, with deduplication across sources.
• Implemented a four-state location eligibility classifier.
• Built a local HTTP application for status tracking and notes with instant persistence.
• Moved all user-specific assumptions into a profile config, so the same engine serves a different user by swapping one file.
• Built a synthetic-data demo mode for safe demonstration.

A first-time founder launching a mobile diagnostic-imaging business needed an operational and digital foundation built from nothing — brand, web presence, client intake, and scheduling.

• Scoped requirements directly with the founder and translated them into a delivery plan.
• Built the website and client intake workflow.
• Integrated Cal.com scheduling into the intake flow.
• Delivered brand and positioning materials alongside the build.

An AI advisory firm that runs its own operations on a multi-agent AI platform designed and built in-house. The platform serves as an operating system for the business: specialized AI agents handle intake, delivery, research, and audit work across two model families, coordinated through a dashboard ecosystem with shared state, persistent memory, and access control.

• Architected a multi-agent environment spanning Anthropic Claude Code and OpenAI Codex CLI runtimes, with per-agent identity, scoped permissions, and audit chains.
• Built an agent memory bridge so agents retain state and prior decisions across sessions.
• Designed and integrated a 7-dashboard ecosystem over webhook orchestration, backed by Supabase (Postgres, Auth, Realtime, Edge Functions, RLS).
• Engineered and continuously tuned prompts and guardrails across a 50+ agent roster for production reliability.
• Implemented cross-model adversarial review — an agent on a different model family audits work the primary agents produce.
• Built human-approval gates on consequential actions and authored the operating doctrine agents work under.
• Deployed and debugged MCP servers and Cloudflare tunnel/DNS configuration in production.

A production voice AI agent that answers inbound calls for service businesses — qualifying callers, capturing leads, routing intake, and booking appointments without a human on the line. It replaces the Tier-1 phone handling a small business otherwise cannot staff.

• Built the agent on Vapi.ai and integrated it with n8n, Gmail OAuth2, and Cal.com for self-service booking and lead capture.
• Designed the conversation state handling and intake routing logic.
• Iterated prompts and interruption/turn-taking behavior across versions to eliminate false interrupts.
• Diagnosed defects from call recordings rather than transcripts, after finding that speech-to-text artifacts masked the real failure modes.
• Built and debugged the webhook layer connecting the agent to downstream scheduling and notification workflows.