Mendy Ezagui

• 96% — Reduction in Average Time to Assign Leads Using Round-robin Routing
• 100% — AI-driven After-hours Lead Engagement and Response Coverage
• 30% — Reduction in Lead Leakage from Existing Contacts
• **50% ** — AI Classification Eliminated Spam and Support Leads from Sales Queues
• 25%+ — Leads Automatically Matched to Accounts to Identify Existing Customers and Product Ownership
• 100% — Leads Scored for Prioritized Sales Engagement

• Manual, inconsistent lead routing caused hours–days assignment delays.
• SDR queues flooded with spam/support leads, hiding real prospects.
• No prioritization against SLAs or first-contact windows.
• Existing customers entered as new leads, causing leakage and poor conversion.
• After-hours inquiries went untouched, reducing engagement and pipeline quality.

• Ran discovery sessions mapping intake to qualification to expose routing and queue bottlenecks.
• Implemented Salesforce round-robin ownership with SLA tracking for rapid assignment.
• Deployed AI after-hours engagement to capture intent and queue leads for follow-up.
• Added AI classification to divert spam/support away from sales queues.
• Enabled lead-to-account matching and Einstein scoring to prioritize outreach and preserve customer context.

• 96% faster lead assignment
• 100% after-hours lead engagement coverage
• 50% reduction in queue noise from spam/support
• 30% reduction in lead leakage
• All leads prioritized via scoring, improving conversion and protecting rep capacity

This is the story of most startups. A solution designed to solve one issue now needs to mature and evolve with a stable foundation to serve many purposes.

Nucleus was a young company and, at first, had specific problems they attempted to solve. With success, the issues quickly evolved to partial-filled orders, international orders, multiple carriers, same-day orders, Amazon Prime orders, shipping rules, to name a few.

At first, we utilized reporting to identify problematic orders. We quickly determined that it’s a waste of time and a hindrance to scaling.

We then designed a backorder engine, and shortly thereafter, an allocation engine. These worked in concert, and from that point forward, we rarely engaged with orders, and lucky for us, our order volume would increase exponentially.

• 30% — Calls Fully Resolved by AI with Zero Human Handoff
• <2 Minutes — Average Handle Time per Call

• Drivers calling a human dispatcher waited 2–3 minutes just for caller identification and route lookup.
• Dispatchers were capacity-bound. Inbound calls were missed entirely.
• No structured capture of call intent or follow-up data — calls disappeared after they ended.
• High volume of repetitive question types (app issues, route issues, pay/financial questions) was consuming dispatch.

• Built a voice agent on Retell with custom prompt architecture and guardrails for the medical/field-service domain.
• Stood up bidirectional integrations with Salesforce CRM, Slack, Google Drive, and Gmail — caller is identified and contextualized within seconds.
• Deployed in two languages (English and Spanish) with fluent, fast conversational delivery.
• Implemented structured intent capture so every call writes a clean record into CRM — no more lost context.
• Designed escalation paths for calls that need a human, with full context handed off.

• Around 30% of inbound calls fully resolved by AI with zero human touch.
• Caller identification and route lookup compressed from 2–3 minutes to seconds.
• Average handle time under 2 minutes; many calls fully resolved in under 1 minute.

• 1 Hour — Time to Spin Up a New Vertical Demo, Reduced from 5 Days
• 10% — Wasted Call Time, Reduced from 40%

• Voice AI is generic by default—buyers don’t see themselves in a horizontal demo.
• Building a tuned, vertical-specific agent usually takes 5+ days of engineering work per vertical.
• The communication layer in most companies is fragmented across voice, chat, CRM, and ticketing—orchestration is custom every time.
• Demos and live agents traditionally require separate setup, slowing time-to-value.

• Designed and built the platform end-to-end as founder—product, architecture, prompt design, and integrations.
• Productized the vertical-demo pipeline: templated agent build on Retell, Supabase-hosted guide pages, and Webflow integration wiring “Demo X Agent” buttons directly into the marketing site.
• Built the prompt architecture and persona system so verticals are configured, not coded.
• Stood up the AGENT_MAP layer, routing visitors from the marketing site directly into the right live demo.
• Managed a 2-person team executing the build cadence.

• Time to spin up a new vertical demo compressed from 5 days to 1 hour.
• First customer deployment (Rapid Medical, field service voice agent) live in production.
• Two live voice agents in production, expanding to additional business units.

• **73+ ** — Tools Auto-generated From OpenAPI Spec
• 105 — HOA Associations Live

A property manager had migrated to a powerful but awkward HOA management platform: its API lacks a name search and an assignee field, blocks dynamic IPs, and forces a clunky dual-auth scheme. A solo operator needed to monitor and act on a portfolio of 100+ associations without living inside the platform’s UI—and wanted AI assistants to query it safely.

I built a spec-driven MCP server (Node.js and TypeScript) that auto-generates ~73 tools from the platform’s OpenAPI spec, then added 6 composites that fill real API gaps—name search, account/association snapshots, and a manager queue whose priority heuristic encodes domain knowledge (active water intrusion outranks fine-accruing violations outranks overdue items). Tiered guardrails keep it read-only by default with destructive ops gated. I exposed it to Claude, ChatGPT, and Codex via OAuth 2.1 (DCR and PKCE), hosted behind a tunnel, plus a Slack bot with a 7 am ‘what’s slipping’ digest.

Live across a 100+ association portfolio. A solo operator can query and triage the whole book from Slack or any AI assistant; a daily digest surfaces what’s slipping before it escalates—a closed platform made AI-queryable.

• ~29 Tables — Unified Operator Data Model
• 20+ Views — CRM, Finance, Projects, AI Memory

As a solo operator, I was juggling deals, tasks, projects, invoices, calendar, contacts, and AI context across disconnected tools—and constantly re-pasting context into AI chats. I needed one system that unified my entire operation and that AI agents could read and write directly, safely, without me babysitting every action.

I built a single-page operator system on React and Supabase (PostgreSQL and RLS, about 29 tables, on Vercel), unifying CRM, deals, financials, projects, an AI store, and a multi-LLM playground across about 20 views. The standout is a remote MCP server (Cloudflare Worker and Durable Objects) that lets AI assistants read and write the workspace via scoped, rate-limited tools, including committing code to the app repo to trigger a deploy. A daily cron runs three parallel AI analyses (pipeline, A/R, and at-risk contacts). Scoped dual tokens, RLS, and confirmation gates keep destructive ops safe.

My daily operating backbone: one source of truth for revenue, work, and AI context, with autonomous morning sweeps surfacing what needs attention—and a closed AI loop where assistants can ship the app they run on.

• 0 Upload — CSV Never Leaves the Browser
• 0-100 — Pipeline Integrity Score

B2B SaaS teams report pipeline numbers they can’t fully trust—overdue, stalled, and zombie deals, missing fields, and quarter-end clustering quietly inflate the forecast. RevOps leaders needed a fast way to see how much of their reported pipeline is real, without handing sensitive CRM data to a vendor.

I built a self-serve diagnostic that scores a pipeline from a CSV entirely in the browser—the data is never uploaded. A hand-written RFC-4180 parser maps Salesforce, HubSpot, and generic CRM columns, and a scoring engine produces a 0–100 pipeline integrity score with bands, flagging overdue, stalled, and zombie deals, missing fields, quarter-end clustering, and duplicates—each quantified in dollars and as a percentage of the pipeline. It generates a branded Excel teardown client-side and captures the lead with its score attached. Identical scoring math is mirrored in standalone Python tools.

A live, zero-liability lead magnet: prospects get an instant, credible read on their pipeline integrity while their data never leaves their machine, and qualified leads arrive with a computed score that frames the sales conversation.

• 6+ Tools — GTM Stack Coordinated in One View
• 30-Day — Pipeline Creep Detection

Series A–C SaaS revenue teams run a sprawl of disconnected tools—CRM, sales engagement, conversation intelligence, enrichment, and data platforms—with no single place to see forecast health, pipeline quality, or what their AI layer is actually doing. There was no tangible way to show CROs an agentic RevOps vision in a sales conversation.

I designed and built a polished, self-contained demo of an agentic RevOps coordination layer: a Signal Intelligence feed, an “Ask AI” natural-language pipeline-query box, Strategic Actions, simulated Agentic Activity (classification, engagement, and BANT agents), a quarterly forecast, pipeline-quality scoring, and 30-day pipeline-creep detection. It’s deployed on Vercel and served at a custom hostname through a hand-written Cloudflare Worker reverse proxy, with OpenGraph and social assets for sharing.

A high-conversion sales asset: prospects click through the full product vision at a branded URL, clearly framed as a demo, then are prompted to build it on their own data, making an abstract agentic-RevOps pitch tangible.

• 13 Tables — Real-time PostgreSQL Schema

Event coordinators juggle guests, seating, timelines, check-in, and last-minute changes across spreadsheets and group chats, while guests have no single place for their seat, schedule, and questions. The need was for a real-time system serving both the event coordinator and the guests attending.

I designed and built a dual-app event platform—separate admin and guest builds—on React and Supabase with a 13-table real-time schema. Coordinators get an auto-advancing timeline, real-time team chat, tier-targeted broadcasts, a voice-input event creator, Claude Vision card-scanning, and a full seating manager with check-in. Guests get passwordless magic-link access to their seat card, schedule, announcements, and an AI concierge. A multi-provider LLM gateway keeps API keys server-side and parses messy guest lists into structured rows.

Taken from concept to a live first customer. Coordinators run the entire event from a single real-time console while guests self-serve everything they need—replacing spreadsheets and group texts with a single, live source of truth.

A consulting practice had outgrown a Squarespace site—no control over structure, lead capture, content, or the ability to host interactive product demos. It needed a fast, self-owned marketing and lead-gen platform that a non-developer could edit and that could host live, interactive tools.

I replaced it with a self-built Next.js 15 (App Router) app: a block-based headless CMS over Supabase with magic-link admin auth, public marketing and insights routes, lead-capture and newsletter APIs, ISR, and generated OG cards, deployed to Cloudflare Workers via OpenNext. The home page is a portfolio grid of the product line; the site also hosts a live diagnostics tool, an interactive demo, and an automated daily Morning Intelligence Brief worker that scans industry voices and emails subscribers.

A fully self-owned platform replacing Squarespace: every page and post is admin-editable, leads flow straight into the back end, and the site doubles as a live showcase for interactive products and an automated content engine.

• **3 ** — Salesforce Agentforce PoCs

Lenders across merchant cash advance, community finance, and cannabis lending needed faster, defensible decisions—yes, no, or human review while a broker is on the phone, investor-ready impact reports without manual wrangling, and compliance packets prepared before a human opens a file. Generic AI chat couldn’t enforce policy or stay the system of record.

I designed and built three Salesforce Agentforce proofs of concept for lending, each pairing deterministic Apex (policy enforcement, Salesforce as system of record) with an Agentforce action that exposes the decision conversationally: a merchant-cash-advance speed-to-decision desk with trigger-based scoring; an impact-reporting workflow comparing baseline commitments to actual outcomes and generating investor narratives without inventing numbers; and a compliance-review workflow modeling license verification, KYC/OFAC, beneficial owners, and SAR signals.

Three reusable, regulated-industry Salesforce AI workflows that deploy Agentforce responsibly: Apex enforces the rules and owns the data while the agent handles the conversation — turning slow manual underwriting into fast, auditable decisions.

An event-rental brand needed a real storefront—branded catalog, rental inquiry flow, policies, and generated rental agreements—built and maintained in code rather than clicked together by hand, so catalog and content changes could be automated and version-controlled.

I built a code-first Shopify storefront: a custom Liquid theme (branded home, looks, and forms sections) plus Node.js automation scripts that drive the Shopify Admin GraphQL API to publish and clean the catalog, build collections and menus, generate FAQ and policy pages, and seed inventory with dimensions and options. A cream-and-gold brand system runs across theme and assets, with a generated-PDF pipeline for delivery, will-call, and sprinter-service rental agreements.

A fully code-managed storefront where catalog, content, and brand are version-controllable and scriptable, replacing manual Shopify admin work with repeatable automation, and ready for launch.