Systems that move from idea to real execution.

Appointment-based businesses often receive repetitive service and booking questions through WhatsApp, including outside working hours.

A multi-modal WhatsApp receptionist for text and voice messages, using Pinecone RAG for clinic-grounded answers and MCP integrations to check calendars, book appointments, and send confirmation emails.

The implementation demonstrates consistent appointment capture, after-hours availability, and controlled escalation without claiming unverified client results.

• Faster response outside normal business hours
• Fewer repetitive front-desk interactions
• Consistent appointment capture with human escalation

Sensitive, uncertain, or exceptional requests are routed to a human instead of being actioned automatically.

Calls, availability checks, reminders, and follow-up create repetitive work and missed booking opportunities.

A Vapi and Retell AI voice agent with ElevenLabs speech, n8n orchestration, calendar booking tools, FAQ handling, and live human transfer for complex conversations.

The build demonstrates how routine booking calls can be handled consistently while conversations requiring judgment transfer to staff.

• Always-available handling for supported call intents
• Calendar actions for standard booking requests
• Human transfer for complex or sensitive conversations

Unsupported requests, uncertainty, and policy exceptions trigger human transfer or a callback task.

Patient inquiries can require different handling for scheduling, general questions, and follow-up, creating a repetitive triage process.

A multi-agent n8n system using DeepSeek and memory, with three specialist agents that classify Gmail inquiries and route them to the correct handling logic.

The technical project demonstrates classification, specialist routing, memory, and an architecture that can expand to additional channels.

• Structured inquiry classification and routing
• Specialist agent handling for different request types
• Architecture designed to extend to more channels

Low-confidence classifications and sensitive messages can be routed to a human review path.

Booking questions and availability checks required manual replies across multiple conversations.

A structured booking system that captures requirements, checks the next action, stores booking details, confirms the request, and escalates exceptions.

Faster responses and a repeatable booking workflow with cleaner operational handoff.

Sales teams need faster personalized drafting without allowing unreviewed AI messages to reach prospects.

An agent that reads structured lead data, drafts the message, collects approval or feedback, revises when needed, and sends only after approval.

Faster personalized drafting while preserving human control over client-facing communication.

Every external message requires explicit human approval before sending.

Prospect research and lead preparation were fragmented across sources, spreadsheets, and manual checks.

A workflow that gathers lead data, validates and enriches records, applies qualification logic, and organizes the result for outreach.

The pipeline demonstrates repeatable research, validation, enrichment, and CRM-ready lead preparation.

• Structured lead research across connected sources
• Validation and enrichment before CRM handoff
• Human review remains available before outreach

Validation rules and review steps prevent unverified records from moving directly into outreach.

Teams lose time searching documents and repeatedly answering questions that already exist in internal material.

A retrieval system with document ingestion, embeddings, vector search, conversation memory, and cited responses.

The assistant demonstrates grounded retrieval, source citations, and explicit handling when approved evidence is missing.

• Retrieval from approved business information
• Citations that let users check the supporting source
• Refusal or escalation when evidence is insufficient

Answers are grounded in retrieved sources, but human review remains appropriate for high-impact decisions.

AI assistants could generate answers but could not reliably interact with real tools, data, or operational systems.

An MCP integration layer exposing controlled tools, validated inputs, structured outputs, and business actions to AI agents.

Agents that can move beyond chat and complete controlled work across connected systems.

Tools use validated inputs, scoped permissions, and structured outputs rather than unrestricted system access.

Restaurant staff repeatedly answer questions about menus, policies, opening hours, availability, and reservations, while unusual or sensitive requests still require personal attention.

An n8n assistant using DeepSeek, PostgreSQL conversation memory, retrieval-augmented generation, a restaurant knowledge base, Gmail routing, and Ollama embeddings. It collects reservation details and sends a structured booking request to the restaurant team.

The demonstration shows faster routine responses, consistent knowledge-grounded answers, organized booking-request collection, and human control for exceptions without claiming automatic reservation confirmation.

• Faster responses to routine restaurant questions
• Fewer repetitive interruptions for staff
• Structured booking requests with human escalation

The team confirms availability and handles uncertain, sensitive, or exceptional requests; the workflow does not claim that a reservation is automatically confirmed.

Teams spend time manually researching prospects, collecting business information, formatting records, and copying inconsistent data into spreadsheets.

An n8n agent that interprets research criteria, uses Google Search, extracts and formats relevant lead information with JavaScript, and appends structured records to Google Sheets for review.

The system reduces repetitive research and spreadsheet entry, produces more consistent lead records, and centralizes prospect information for human review and follow-up.

• Less manual prospect research and data entry
• More consistent review-ready lead records
• Centralized research and status tracking

Records remain review-ready rather than automatically verified or converted; the case study does not claim validated emails, autonomous outreach, CRM updates, or customer conversion.

Service teams lose time answering repeated WhatsApp questions, listening to voice notes, checking images, searching business information, and transferring requests into other tools.

A multimodal n8n workflow that prepares text, audio, and image inputs before routing them to a DeepSeek agent with conversation memory, a knowledge base, web search, and an MCP client. It can return text or audio responses and preserve a human handoff path.

The project demonstrates one controlled system for common WhatsApp message formats, approved tool access, consistent information retrieval, and staff involvement when judgment is required.

• Text, voice-note, and image handling in one workflow
• Approved knowledge and MCP tool access
• Text or audio responses with an exception path

External tools are scoped and approved, and uncertain or sensitive requests remain available for staff review. The conversation image is an anonymized example, not a client-result claim.