Holti Bitri · CTO · Milan AI agents and people, one team. I design it.

I design and lead workflows where AI agents and people build software together. Agents get written rules and a check at every step. People keep the architecture and the decisions.

I have been writing software since 2008: .NET and Oracle developer at Monte dei Paschi di Siena, then systems integration, architecture, and leading an innovation unit. Today I am CTO at TIG Factory.

My job has changed shape. Most of the code is now written by AI agents. I decide the architecture, write the rules they work by, set the checks, and answer for what goes to production.

This is not an experiment. It is how I build a trading platform, a podcast that ships every morning without me, and the tools my colleagues at TIG use every day.

Before that I worked inside PwC, Medtronic Italia, Reply (for Eurizon) and Magneti Marelli, spent four years at Nexive and eight at Cloudia Research, from architect to Innovation Director. That is where I learned what software needs to hold up in a company, and it is what I now ask of the agents.

Microsoft certified since 2011 (MCSD, MCTS), Anthropic certified since 2026 (Claude Code, Claude Cowork).

Holti Bitri, CTO at TIG Factory
18
Years of software in production
100+
AI FinCast episodes shipped with no human intervention
Worked with
  • PwC
  • Medtronic Italia
  • Reply / Eurizon
  • Magneti Marelli
  • Monte dei Paschi di Siena
  • Nexive

Five rules,
learned from things that run.

  1. 01

    First, decide whether AI is needed

    A client asked me for AI. Thirty minutes in, it was clear he needed an if statement. Being able to say so is part of the job.

    Read the post →
  2. 02

    Agents get a contract, not a prompt

    Role, rules, sources, what it must never do. Written down, versioned, corrected after every mistake. At TIG Factory the company knowledge base is kept by an agent with one rule at the top: never make anything up.

    Read the guide →
  3. 03

    Sensitive data never reaches the model

    At TIG, a check on the production database used to take half an hour by hand. Now it takes seconds: personal data is masked on the PC and the model only gets the summary.

    Another case: meetings transcribed locally →
  4. 04

    Every step has a check

    An agent that writes without verification produces fast code and slow problems. Tests, review and acceptance live inside the workflow, not after it.

  5. 05

    Decisions stay with people

    Agents propose, write and verify. What to build, on which architecture and when it ships is decided by whoever answers for it.

The method,
put to the test.

Open Source · Personal

AI Meeting Assistant

Fully local AI pipeline for transcribing meetings, identifying speakers, and generating executive summaries. Works with any platform: Teams, Zoom, Webex. No cloud, no subscriptions, no data leaving your machine.

PythonFastAPIDockerWhisperOllama
View on GitHub →
Product · Cloudia Research

AI4Teams

An AI-powered hub designed to enhance team collaboration, alignment, and productivity within the Microsoft 365 ecosystem.

Microsoft 365Azure AIPower Platform
Product · Cloudia Research

AI4Orchestration

Intelligent orchestration platform on the Microsoft stack, enabling enterprise-grade workflow automation and AI-driven process management at scale.

AzurePower AutomateAI Orchestration

Eighteen years of real software,
before the agents.

2026 → today
CTO
TIG Factory
I lead technology strategy and delivery at TIG Factory, the business unit of TIG (The Innovation Group) focused on custom digital solutions and enterprise software. I bring AI agents into how the team designs, writes and verifies software.
Technology StrategyAI in Engineering.NETAzure
2024 → 2026
Innovation Director
Cloudia Research
Directed the Microsoft Hub (18+ consultants), overseeing project execution across Azure, Custom Dev, and Power Platform. Led product R&D for internal SaaS solutions and acted as Solution Architect for enterprise clients.
Team LeadershipSolution ArchitectureProduct R&DPre-sales
2018 → 2024
Solution Architect · Team Leader
Cloudia Research
Designed and implemented cloud-first architectures on .NET Core and Azure. Application modernization, workflow automation, and data-driven reporting for enterprise clients. From 2023, coordinated .NET development and the integrations between CRM, ERP and PLM systems.
.NET CoreAzureCloud ArchitectureDevOps
2013 → 2017
Integration Specialist
Nexive
Architected QArc, a custom SOA middleware layer integrating Oracle, DB2, and Salesforce ERP across enterprise environments. Contributed to Nexibus, an Enterprise Service Bus built on TIBCO.
SOA MiddlewareTIBCOSystem IntegrationC#
2011 → 2013
.NET Consultant · Senior Developer
WICE Group
Consulting at PwC (data access layer over SAP, Active Directory and FIM, onboarding portal on SharePoint), Medtronic Italia (internal applications, multilayer architecture, training the IT team), Reply for Eurizon (services for real-time operations on insurance policies) and Magneti Marelli (power-of-attorney portal on SharePoint).
PwCMedtronic ItaliaReply · EurizonMagneti Marelli
2008 → 2011
Senior .NET Developer · Oracle DWH Developer
Monte dei Paschi di Siena
Owned ETL pipelines and metadata management for the financial data warehouse using Oracle Warehouse Builder. Delivered web reporting and banking services across Anagrafe and Finance divisions.
Oracle OWBETL / PL-SQLReporting ServicesC# / ASP.NET

Guides and notes
from the work.

Illustration: a grid of grey dots with fifteen blue dots, connected by lines to a column of fifteen dots that converge into an audio waveform; at the top the time 07:00 and an on air sign

An AI-generated podcast every morning, without me: several agents at work, each with the right model for its task

A daily podcast written, voiced and published by a team of AI agents. How the work is split, why each role has a different model, and the two mistakes that taught me the most. Hosted here, in full.

Illustration: a morning briefing showing status OK and five emails on the list, two of them drawn with dashed outlines and marked as not found

My AI assistant was making up emails and I didn't notice: why AI agents make up data, and how to catch it in time

For days my morning briefing delivered false data without a single error. The three failures underneath, why no check caught it, and the six rules that came out of it. Hosted here, in full.

Peugeot 3008 infotainment screen before and after the NAC firmware update: version 42.03.31.32 from 2020 on the left, version 44.07.33.32 on the right

One error, four causes: updating the Peugeot 3008 (2nd gen) firmware without losing your mind

The official tool shows one message for four different problems. The full USB procedure for the NAC Wave 4 firmware (Peugeot, Citroën, DS, Opel): UIN, UpdateID, license, CRL, folder structure. Hosted here, in full.

Illustration: a desk in a night-time office in front of a city skyline, with a holographic knowledge graph labelled Second Brain and panels showing an AI agent linking nodes, summarising the day's notes and suggesting related content

An LLM wiki with Obsidian: building a second brain the AI reads before it answers

Every chat with a model starts from nothing: who you are, what you are building, what you already tried. One rule about structure, a written contract instead of a prompt, and a lint that stops the vault from rotting. With the technical backstage, hosted here.

Illustration of a desk setup: a monitor showing the meeting dashboard, a laptop on a video call with the team and a tablet showing the summary, with labels for automatic transcription, chapters and task allocation

Transcribe and analyse meetings locally, without sending the data to the cloud

Copilot needs three conditions at once, and real meetings rarely meet them. The condition that actually matters is the fourth one: may that data leave the company at all? Two containers, Whisper and Pyannote on CPU, a local LLM. Full architecture, hosted here.

Let's build
something together.

If you want to bring AI agents into how your team builds software, or you have a project that needs to reach production, write to me. I answer personally.