The AI systems I have built for NexaVision AI

When I talk about NexaVision AI, I do not think only about a website or a collection of isolated automations. For me it is a real product lab: a place where I am turning repetitive business problems into AI systems that can run, be measured and be maintained.

In the internal workflow repository I currently have more than 30 n8n exports, several lightweight HTML frontends and a documented AI phone call system with VAPI. Some are products ready to adapt to clients, others are functional templates or advanced prototypes. The important part is that they all share the same idea: AI only creates value when it is connected to real tools, real data and real business decisions.

I already explained the broader company vision in NexaVision AI: what I am building and why it is one of my most important projects. In this post I want to go one layer deeper and organize the concrete systems I have built.

The common architecture

Even though every workflow solves a different use case, most of them follow a similar architecture:

1. Input: webhook, form, Gmail, Telegram, WhatsApp, Google Sheets, RSS, phone call or event.
2. Context: client data, history, configuration, documents, calendar or lead database.
3. AI model: Gemini, OpenAI or another LLM to classify, write, decide or summarize.
4. Action: send emails, publish content, create events, update sheets, notify through Telegram or escalate to a person.
5. Traceability: store the result, log errors and leave an output that can be reviewed.

That pattern matters more than the specific node. A good AI system is not "a prompt": it is a chain of decisions with limits, fallback and verifiable output.

Customer support agents

One of the most important families is conversational agents. I have worked on several channels:

• Web chatbots for WordPress and Hostinger, with automatic answers and human escalation.
• Gmail agent, designed to answer or classify incoming requests with context.
• WhatsApp agent, both for text messages and voice notes.
• AI phone receptionist, connected with VAPI and n8n webhooks to check, create, update or cancel appointments.

The hard part here is not answering in natural language. The hard part is knowing when not to answer automatically, when to ask for more information and when to move the conversation to a human. That is why these systems are designed with escalation, context and interaction logs.

Lead generation and follow-up

I have also built several systems to automate commercial acquisition:

• Automatic lead generation, using Google Places/Maps and AI enrichment.
• Weekly lead search with SerpApi, to find opportunities on a recurring schedule.
• Lead scraper webhook, to launch searches on demand from a frontend or external system.
• Lead tracker, to follow the commercial status of each opportunity.
• Google Maps email scraper, focused on building commercial databases.
• SQL lead summary, to turn raw lead data into an executive view.

In these workflows, AI does not replace the salesperson. It removes noise: scoring leads, summarizing information, generating first messages and prioritizing who deserves attention first. For a small company, that can be the difference between having an infinite contact list and having an actionable pipeline of opportunities.

Content, marketing and social media

Another line of work has been turning ideas or external sources into publishable content:

• YouTube to WordPress, transforming videos into articles through transcription and writing.
• Automatic blog system, split into RSS ingestion, SEO rewrite with images and WordPress publishing.
• Multi-channel marketing campaign planner, for weekly content planning in different formats.
• Social media manager from Telegram, where a user sends a photo or description and the system generates copy for social channels.
• Dealership social manager, adapted to a vertical use case.
• Advertising videos with Vertex AI, focused on creating visual pieces from prompts and product data.

This block is especially interesting to me because it mixes automation with editorial judgment. I do not want systems that publish content blindly. I prefer flows where AI accelerates drafts, structures and adaptations, while leaving room for human review.

Internal operations and business automation

NexaVision AI has also helped me build systems that are less visible from the outside, but can save a lot of hours inside a company:

• Booking management with Google Calendar, including a web widget, availability checks, confirmations and a restaurant-specific case with WhatsApp, AI and n8n.
• Invoices through Telegram, with document processing and storage in Google Drive/Sheets.
• Automated ecommerce deployment, designed to prepare deliverables or configurations when a client buys a product.
• Smart form, designed as a structured entry point to qualify requests.
• AI mass email, to generate and send personalized campaigns with review.

Here the key is connecting existing pieces: calendars, spreadsheets, email, forms, Drive folders and dashboards. Many companies do not need a huge platform built from scratch; they need their tools to talk to each other.

HR, verification and analytics

I have also developed more specialized workflows:

• AI CV filter, to analyze candidates against concrete requirements.
• AI-generated text detector, combining heuristics and model evaluation.
• KPI insights, to summarize indicators and turn them into an executive reading.
• Idea scoring through Telegram, to register ideas and prioritize them with criteria.
• Crypto influencer scraper, using web search and an AI agent to research profiles in the sector.

These systems show another side of AI: not only generating text, but helping with decisions. Classifying, comparing, detecting patterns and reducing the cost of reviewing information.

What I have learned building these systems

After building so many workflows, some lessons keep repeating:

• The prompt is not the product. The product is the full experience: input, data, model, output, review and maintenance.
• n8n becomes very powerful when combined with custom code. Visual nodes speed things up, but real cases usually need JavaScript, APIs, validations and error handling.
• AI needs clean context. If the input data is messy, the model improvises.
• Human escalation is part of the system. Automation does not mean removing people, but reserving them for the important decisions.
• Traceability creates trust. If a workflow fails or makes a decision, it must be reviewable.

Where I am taking it

My goal is not to have a folder full of JSON files. My goal is to turn these systems into clear packages: customer support, lead generation, content, bookings, HR, analytics and internal operations.

Each package should be adaptable to a client with their credentials, tone, data and rules. That is the difference between a nice demo and an automation that can live inside a company.

You can learn more about the project at nexavisionai.net, read the NexaVision AI project page in my portfolio or reach out through the contact page if you want to talk about agents, n8n or AI systems applied to business.