Every system I have built started with a number the business wanted to move: tickets per day, leads per month, sites per month, minutes to the first reply. The architecture follows the number, not the other way round.

Prompt, retrieval, orchestration, integrations, servers, security, documentation. If it runs in production, it is mine until it is monitored, tested and handed over with the documents the next person needs.

Langfuse evaluations on retrieval and answers, CSAT, a comparison of every new flow version with the previous one before release, and a weekly improvement loop on the real conversations. A RAG score that went from below 0.6 to 0.9 is the kind of proof I look for.

Before building, I look at the best product in the category, Intercom for the support widget, Zendesk for the CRM, and I write down what beating it means for this client. Then the blocks, the milestones and the QA tests follow that list.

About 50 support agents handled roughly 4,000 tickets a day by hand, with peaks of more than a hundred conversations in five to ten minutes. Under overload the first reply took up to 18 minutes and customer satisfaction sat at 2.5 out of 5. Every ticket, including the trivial ones, cost human time, and the knowledge needed to answer was scattered across the CRM and years of message history. The brief: a widget that answers most tickets by itself, in the customer's language, and leaves humans only the cases that need judgment. The benchmark I set before building was Intercom.

A chat widget on the client's site, deployed through Google Tag Manager, that understands the request in any language, checks the customer's account when they are logged in, answers from a knowledge base built on the client's own CRM and message history, and hands complex cases to a human in Zendesk with the whole transcript and account context attached. n8n orchestrates the main, escalation and reporting flows; LangGraph runs the cyclic logic per intent, so the graph decides step by step whether to answer, ask, verify or escalate. Each agent has its own prompt, GPT-4o mini as main model and Gemini 2.5 through OpenRouter as fallback. Pinecone holds the knowledge base, MongoDB the shared memory, Redis the cache. I set up and hardened the servers myself: everything runs in Docker on the project's own infrastructure, with monitoring on the servers and on every n8n node so that bugs surface before the business sees them.

• One prompt, one main model and one fallback model per agent. The widget itself has a fallback version for when the AI path is down.
• GPT-4o mini for cost and latency: average reply ~3 s, long conversations capped around 7 s, peak 9 s.
• Shared memory in MongoDB feeds a weekly improvement loop on real traffic: answer quality, prompts, CSAT, LLM security.
• Every new version of a flow is compared with the previous one before release, node by node, so behaviour changes are caught before customers see them.
• Vision and OCR: a customer sends a photo, a receipt for example; the data is extracted, stored in MongoDB and used by the agents.

• Servers built and hardened from scratch; the data path between orchestrator, server and browser is protected end to end.
• Penetration tests run through agents I built with rules based on the permission model, grey-box tests on API, latency and infrastructure entry points, review by the internal security team and a black-box test by an external company. Every finding fixed by me.
• Prompt-injection defence and control of what reaches the model; sensitive data never leaves the perimeter without a reason.

• Solo project, more than six months to release, divided into blocks: each block shipped with its own QA, monitoring and documentation.
• All automations versioned on GitHub with CI/CD pipelines.
• Final deliverables: API documentation, a playbook for the customer-service team and the technical project document.

The 50-person support team was not cut. It was trained on the new flow and partly offered to other companies as a service, so the business turned a cost centre into a new line of revenue.

Migration to Qwen 3 from Hugging Face with Qdrant, built with the ML and DevOps teams, for privacy, latency, control and the option to improve the model. The local model also acts as a filter on the data sent to external LLMs. Fine-tuning in preparation, then the same model will serve the internal teams and the widget itself.

Brands needed dozens of websites every month to grow their presence and send traffic to the main page: satellite sites around one or two brands, each specific enough to be indexed and read. There was no content team to research, write, design, illustrate and build them, and each site had to be specific enough to be indexed and cheap enough to make dozens a month.

A pipeline of Hermes agents, self-hosted on the company server, with a dedicated agent for research, one for scraping and composition, one for design, and a final script that publishes. Two ways in. From an example URL: the agents research and scrape a similar site and compose a new one inspired by it. From scratch: the request enters through MCP from Claude Code or Codex, research runs on Perplexity and the Anthropic API, the system scrapes what the operator asked for, then runs a short interview to collect the facts only a human knows, and builds the site. GPT-6 Astra generates the images. One click clones a site with a slightly different design, or a completely different one. A database of every run grows with use and makes the second mode better over time. Output is static HTML, exported and deployed on the client's servers through the DevOps script.

Templates give the same site every time, and search engines know it. Agents research the topic, interview the operator and write for that topic, so each site is specific, readable and indexable. Cost stays near five euros because the expensive steps, research and images, are cached in the run database and reused.

• Keyword research and keyword-to-page mapping; title tags and meta descriptions; H1, H2, H3 hierarchy.
• Clean URLs, canonical tags, sitemap.xml, robots.txt; internal linking.
• Alt text and light, optimised images for crawling; Core Web Vitals; mobile first.
• Structured data in schema.org JSON-LD; Open Graph.
• GEO, Generative Engine Optimization: FAQ sections and llms.txt so that AI search engines can read and cite the site; hreflang for multilingual sites.

About a hundred sites a month, five to ten minutes each, around five euros per site. The sites are satellites of one or two brands and send traffic to the brand's main page; traffic is tracked at brand level.

Support, sales and management worked on separate tools and saw the same customer from different angles. Business events, fraud signals, payments and refunds, blocked accounts, affiliate and influencer activity, reached people late, through reports, instead of the moment they happened. The brief: one platform for all three teams, with Zendesk as the benchmark to beat, built on the company's own servers.

One platform for support, sales and management: the same customers, real-time alerts, voice, SMS and Telegram from one screen, a knowledge base in the style of Notion embedded in the product, analytics on AWS and Tableau, and an AI layer that routes requests and answers from the knowledge base. Architecture defined together with DevOps, the development department and the client's PM and Quality Manager; Kubernetes microservices on company servers, one service per engine, front end entirely in TypeScript with parts in Java. I lead three AI engineers: two-week sprints with a reserve of about ten hours for urgent fixes, QA test at the end of each milestone against the benchmark, async improvements, blocker management, daily problem solving with Customer Service, the Quality Manager and the QA manager.

• Backend and front end.
• Access management: roles and permissions.
• Workspace for the support operators.
• Analytics connected to AWS and a Tableau workspace.
• Business alerts embedded in the CRM from department data on AWS.
• Real-time alerts over Kafka, built with an ML engineer.
• Embedded knowledge base, voice and SMS through Voiso, Telegram for customer communication.

Blocks with milestones; a QA test at the end of each milestone and a comparison with the benchmark; improvements handled async; blockers removed daily. Two-week sprints per engineer, Agile mindset, and a small reserve in every sprint for what breaks.

An agent with intelligent routing, a knowledge base with semantic search, and a browser extension of the agent that executes routine tasks for the operators. The CRM serves customer service, management and sales, so the same agent works on support requests and on lead production.

An on-premise LLM (Qwen 3) with full access to CRM data and the internal customer-data platform; an agent with its own knowledge base that helps support answer faster; the browser extension through which support sets actions that run automatically for the customer.

Employees asked HR, Finance and Legal the same routine questions every day: vacation days, expense limits, which template to use. Each answer took time from the people who had to give it, and the knowledge lived in documents nobody opened. On the company website, visitors had no self-service and no simple way to leave a request. The company already lived in Microsoft 365, so the answer had to live there too.

One assistant per department, in the Teams channel where people already work. The HR agent holds all the shared HR knowledge, answers the frequent questions and logs everything; Finance and Legal follow the same logic. Agents built in Microsoft Copilot Studio, grounded on SharePoint documents, with Azure OpenAI for the models and Databricks for logs, data sharing and continuous improvement. The same stack powers an assistant on the company website: general information, connected directly to SharePoint and Databricks, with a lead step that asks the visitor to send the request by email when the conversation shows intent. Built right after the support widget, alone, for an internal stakeholder of Acrh Software.

The client already lived in Microsoft 365: Teams, SharePoint, Azure. Building where people already are removed adoption friction; nobody had to learn a new tool. Grounding on SharePoint keeps every answer sourced and auditable, and Databricks keeps the logs that drive the improvement loop.

• HR: vacation, policies, onboarding questions, with the full shared HR knowledge behind it.
• Finance: expense rules, limits, procedures.
• Legal: templates, versions, when a clause needs human review.
• Every conversation logged, so the departments see what people actually ask and fix the documents, not just the answers.

The assistant answers first and asks for the request by email only when the visitor shows buying intent. Leads reach sales with the conversation attached, so the first human reply already knows the context.

Leads arrived by email and were read late. Five sales managers spent their day on routine qualification, copying names and phone numbers, and lost buyers who had already moved on to the next agency. The client wanted a visible innovation on the website and, above all, leads that arrive clean and get answered immediately.

A conversational widget embedded in the client's site with all the information on the properties and buildings for sale. The bot's code and prompt live on Replit, OpenAI is the model, the conversation is designed in Voiceflow, Airtable is the CRM, Make runs the automations. The widget collects the buyer's name, phone and email, the key field. Flow one: the moment a lead is saved, the manager gets a Telegram notification with the lead. Flow two: the buyer receives a personalised email with the availability that matches their request and a button to book an appointment. Flow three: the click opens a request on the manager's Google Calendar and Meet, the manager accepts, the buyer gets the confirmation. Proactive messages fire when the visitor clicks key buttons, which lifted leads a further five points.

Email was the key field: the follow-up with matching homes did the selling while the manager was busy. Proactive messages on key clicks, mortgage calculator, floor plan, added five points of leads on their own. And the managers stopped copying data by hand, because it arrived already structured.

• Content and auto-posting automations on Facebook, Instagram, VK and X via Make, for the agency and for clients.
• The whole way of working of the agency rebuilt in ClickUp as the team grew from 50 to 200 people: HR integrations with automatic onboarding and offboarding, analytics from the client CRM inside ClickUp, structured conversations with roles and permissions, a task manager and a custom scheduling and meeting platform.
• More than 20 client websites built end to end in JavaScript, Next.js and Astro, with API connections to the CRM, deployed on the clients' servers.

Clients needed to answer customers at any hour and qualify every contact without hiring staff. Conversations spiked with every ad campaign, and a missed message was a lost customer. It was the most requested build at the agency, and every client had a different stakeholder, a different channel and a different definition of a good lead.

A bot that starts from a keyword on WhatsApp, Instagram or Telegram, runs a defined flow with an LLM inside, qualifies the contact with a precise prompt written for that client, and files it in a custom CRM built in Google Apps Script, where the lead moves from cold to warm to hot on completed steps. The bots served both customer service, talking directly with the end user, and support for the client's own team. From 2022 the same clients got voice agents: a voice receptionist and an HR agent that collects candidate data, built on Twilio and ElevenLabs; outbound agents on Vapi and Retell that call leads from the client's lists and ask the qualifying questions. Orchestration in Make and n8n with custom code; models from OpenAI, Anthropic and OpenRouter.

• One prompt per client, with the qualification criteria written as explicit questions, never left to the model's judgment.
• Pipeline stages change only on completed steps: a lead cannot become hot without contact details.
• Every bot runs 24/7; when a campaign spikes, conversations grow without a channel going uncovered.

• Voice receptionist on Twilio and ElevenLabs: answers, routes, takes the message.
• HR voice agent: collects candidate data in a conversation that follows the prompt.
• Outbound agents on Vapi and Retell: call the leads, ask the qualifying questions, hand the result to the CRM.
• A Sales agent, and a Finance agent that processes internal data, on the same orchestration.

Automotive group in Rome, more than 1,000 employees, a fleet of about 200,000 vehicles. Power BI KPI dashboards built on SQL, AWS extracts and internal APIs with the IT department, for the whole sales department and for management with alerts. Idle stock down 25% because data moved faster and used cars sold sooner. KPI and invoice control for a supplier worth two to three million euros a year. Four new analysts onboarded and trained.