Selected Work

24/7 bilingual (EN/ES) real-estate voice AI with warm broker transfer

We built a self-managed real-estate voice-AI phone system for a 155-acre commercial development: two named agents (Sofia + Marco), warm-transfer broker bridge, CRM sync, a drip sequence, and a call-analytics pipeline.

• Vanity number 855-917-LAND on 14 physical signs across the development
• Bilingual EN/ES with two named agents (Sofia + Marco)
• Drip sequence spanning Day 1 → Day 90
• 154 commits across the monorepo

• 24/7 bilingual (English/Spanish) AI phone agents with warm-transfer bridge to a live broker
• Twenty CRM sync + SQLite call-analytics database with full transcripts
• Multi-tier email drip automation triggered by call outcome
• Astro marketing site + React 19 + Recharts analytics dashboard for the broker
• Production deployment: webhook server on macOS LaunchAgent → Cloudflare Tunnel

Source page is inconsistent on the drip: chips say 6-tier, body copy says 5-tier.

TypeScript · Bun · Retell AI SDK · Twenty CRM (REST) · SQLite (WAL) · Claude Haiku/Sonnet (Anthropic SDK) · Twilio SMS · Astro 5 · React 19 · Recharts

Outbound voice-AI sales system with a live 3D-orb demo

We built a full outbound voice-AI sales system for LegalShield associates. AI agent “Alex” on Vapi, a browser demo with a real-time 3D audio orb, and post-call analysis.

• 78% of buyers choose first responder (cited in pitch)
• $3,500 engagement price for Phases 1–3
• LegalShield: 4M+ customers, 4.6/5 Trustpilot, 50-state coverage since 1972

• Live interactive voice demo: lead submits form, speaks to AI agent through a morphing 3D orb reacting to real-time audio volume
• Outbound AI voice agent “Alex” built on Vapi with behavioral psychology techniques
• State compliance logic (blocks SC, NJ, AK; routes to information-only flow)
• Post-call structured analysis: lead temperature, objections, outcome classification stored in Redis
• Outbound sync script (sync-to-vapi.ts) deploying updated prompts programmatically

React 19 · TypeScript · Vite · Three.js (custom GLSL shaders + simplex-noise orb) · GSAP · Vapi AI Web SDK · Upstash Redis · Vercel · Bun · Tailwind CSS v4

An AI voice receptionist wired into CRM, shipped as a win-the-deal demo on a live phone line

An inbound voice receptionist for a roofing contractor that answers calls, books appointments, and handles FAQs via Vapi, wired into GoHighLevel for lead capture and pipeline. Latency-engineered (pre-warmed HTTPS agent, prompt caching) for a ~2.8s round-trip, and provisioned on a live demo line to win the build engagement.

• 126 commits
• Live Vapi demo line (Atlanta metro)
• ~2.8s round-trip to CRM
• EN primary · ES/HI ready

• Inbound voice receptionist: call handling, appointment booking, FAQ via Vapi telephony
• GoHighLevel v2 integration for lead capture and pipeline management
• Latency-optimized webhook architecture (Bun + Hono, pre-warmed agent, FAQ prompt caching)
• GSD-tracked 6-phase build roadmap shipped to a hard demo deadline

Bun · Hono · TypeScript · Vapi · GoHighLevel v2 · Railway

A voice agent that takes restaurant orders by phone in three languages

A Vapi voice receptionist for a halal burger / BBQ restaurant that takes orders by phone in English, Spanish, and Hindi. The full menu knowledge base is inlined for low latency (sourced from a 128-dish ordering API), with an order-narrowing flow, catering routing, and verbal escalation to a team member.

• EN / ES / HI dynamic switching
• 128-dish menu knowledge base
• GPT-5.4-nano + Deepgram Flux
• Live demo line

• Multilingual inbound voice agent with per-caller language switching
• Menu KB inlined in the system prompt to minimize latency (128 verified dishes)
• Order flow: narrow by type/taste → recommend → confirm items, mods, spice level → ready time
• Catering / large-order routing + verbal escalation to staff

Vapi · OpenAI GPT-5.4-nano · Deepgram Flux · Peblla menu API

2026 reference-grade B2B services site, Lighthouse CI ≥ 0.95

We built a polished Astro 5 marketing site for an AI-first IT consulting firm as a reference implementation for SEO/AEO/GEO, gating every Lighthouse category above 0.95.

• Lighthouse CI: all categories ≥ 0.95
• 45 commits
• 6 E2E viewports
• 3D parallax hero with procedural canvas layer

• Single content source (src/data/site.ts) feeds all pages AND machine-readable SEO endpoints
• 3D parallax hero using @react-three/fiber with a procedural halide-topo canvas layer
• GSAP + Motion animations including animated founder team cards and a count-up stats band
• Full SEO/AEO/GEO: per-page OG images via @vercel/og, Organization + FAQPage JSON-LD, llms.txt, IndexNow
• Lighthouse CI gate asserting all categories ≥ 0.95 across homepage + 2 legal pages

Astro 5 · React 19 · Three.js / @react-three/fiber · GSAP + Motion · Tailwind CSS 4 · TypeScript 5 · Bun · Vercel · Playwright E2E · Lighthouse CI · Resend · Zod · @vercel/og

End-to-end content strategy + paste-ready rewrite of all 19 pages

We delivered an end-to-end content strategy and paste-ready HTML rewrite of all 19 pages on an aviation MRO leadership-development consultancy’s site, with custom SQL extraction tooling.

• 19 pages fully rewritten and marked ready-to-paste
• 3 blog articles rewritten with aviation-authority positioning
• 26 Markdown deliverable files
• 2 WordPress backup snapshots used as source data

• End-to-end content strategy + paste-ready HTML rewrite for all 19 pages
• Custom Node.js SQL extraction scripts pulling content straight from WordPress backups
• Aviation-authority positioning across 3 rewritten blog articles
• 26-file Markdown deliverable corpus for clean hand-off

WordPress 6.9 · Elementor Pro · Astra theme · Node.js (custom SQL extraction scripts) · Markdown-based copy system

Brand site for an early-stage emotional-reflection app

We built the full marketing site for Clarify™, an early-stage emotional-reflection app: book-launch capture, investor outreach, a press/media hub, and downloadable brand assets.

• Static-export Next.js 16 brand site
• Book launch + investor outreach + press hub
• Spline 3D brand moments

• Full marketing site with book-launch email capture and investor-outreach flow
• Press/media hub with downloadable brand-asset kit
• Spline 3D brand visuals with GSAP + Framer Motion
• Static export for fast, cacheable delivery

Next.js 16 (App Router, static export) · React 19 · TypeScript · Tailwind CSS v4 · Radix UI · Framer Motion · GSAP · Spline 3D · Zod · Playwright

Pay & contract intelligence platform for nurse anesthetists

We built a community-driven pay and contract-intelligence platform for Certified Registered Nurse Anesthetists: an animated landing site with a multi-section data-submission form and Airtable backend.

• Multi-section salary/contract data-submission form
• Airtable backend
• Animated React 19 + Motion UI

• Community-driven pay and contract intelligence for CRNAs
• Multi-section data-submission form with Zod validation
• Airtable API backend for crowdsourced records
• Animated UI with Motion + shadcn/ui

Astro 5 · React 19 · TypeScript · Tailwind CSS 4 · shadcn/ui · Motion · Zod · Airtable API · Vercel · Vercel Analytics

The receptionist metric that catches callers who hang up in the first 5 seconds — and the workflow that texts them back

“Calls answered” hides the callers who left before the greeting landed. We defined the early drop once — a connected call that ended in 5 seconds or less — in a single shared module, surfaced it on every screen that reports calls, and turned it into a workflow trigger so a drop is chased while the phone is still in the caller’s hand.

• One definition (≤5s, connected calls only) shared by four screens
• Warn at 10%, critical at 20% — silent under 5 connected calls
• The rate is clickable: it filters the call log to the calls behind it
• 24-check regression suite, green end to end

• Early-drop rate on the service Analytics tab, with a 30-day drop line over call volume
• Per-location drop column so a multi-site operator sees which door is losing callers
• Agency roster column plus an attention rule — a desk dropping ≥20% sorts to the top of the book
• New call_dropped workflow trigger fires on hang-up, so a desk can text back automatically
• Fixed the data bug underneath it: Number(x) || null was erasing 0-second hang-ups, the sharpest drops of all

Node.js · mastercontroller · masterrecord (SQLite/Postgres) · Twilio Voice webhooks · Next.js 16 · React 19 · Tailwind CSS

A call nobody reached becomes a ticket and a follow-up, not a lost customer

When a number rings out or a caller leaves a voicemail, the desk opens a helpdesk ticket from the call and fires call_missed / call_voicemail into the project’s workflows — what happens next (a text, a WhatsApp message, an email) is the client’s decision, not the platform’s.

• Voicemail only counts as missed once the recording lands
• Ticket carries the transcript and the number
• Hooks never break the call — failures are logged, not thrown

• Helpdesk ticket created from the call, with reason (after-hours / missed / voicemail)
• Workflow events let each client choose their own follow-up channel and copy
• Missed rate feeds the same answer-rate and roster views as the drop rate

Node.js · mastercontroller · Twilio Voice · WhatsApp Cloud API · Twilio SMS

Every finished call updates the board and writes its own summary

A completed call moves (or creates) the caller’s card on the project pipeline, generates an AI summary against the persona’s analysis settings, delivers a signed webhook to whatever the client runs, and fires call_ended for any workflow listening — all off the Twilio callback path, so the webhook still returns fast.

• Post-call work runs off the webhook path
• Metered voice minutes at the workspace’s plan rate
• Transcript lives on the shared conversation spine

• Pipeline card maintained by the AI — no one re-types the call into the CRM
• Post-call summary generated when the persona has analysis enabled
• Outbound webhook delivery of the call envelope + summary
• Usage metering per connected minute for transparent per-call billing

Node.js · mastercontroller · masterrecord · Twilio Voice · Anthropic / OpenAI models · Stripe metering