Cases: installers, production, retail and security

Why it matters: an installation company gets paid only once someone rebuilds what was done. Here the technician describes the job by voice, the customer signs on the phone and the invoice starts from the signed report: no paper to copy out in the evening, no forgotten extra work.

Before: paper work reports retyped in the evening, signatures to chase, materials disappearing from the vans, late invoices because first someone had to rebuild what had been done.

What was built: iTER7 with a job calendar, work reports dictated by voice on the phone, the customer's signature on screen, per-van stock with QR codes, branded PDFs and the electronic invoice from the signed report.

What changed: the work report is ready when the job ends, signed, and the invoice starts from there. No sheets to retype. Before-and-after times are being measured with the client: the numbers will appear here once measured, not estimated.

The AI inside: an assistant you talk or write to in Italian ("my jobs this week", "how many of this item are in van 2", "close the job"): it searches customers, jobs, stock and bills of materials on its own and prepares the action. The dictated work report becomes a draft with materials taken from the catalogue, and the system flags materials used beyond the quote. Quotes can be dictated too. Customers' WhatsApp messages and emails arrive already sorted: what they ask for, which job they refer to, what to do next. Photographed or PDF documents read themselves.

AI security: the AI proposes, a person confirms, and who and when is recorded. The assistant can only perform actions from a closed list, with a cap on proposals every ten minutes; manipulation attempts (prompt injection) are filtered; emails, IBANs, tax codes and card numbers never reach the logs. Coverage of the OWASP risks for LLM applications is documented item by item, including the ones not covered.

Data and AI: a dedicated instance for the company, data exportable at any time. AI models are chosen and changed in configuration (the interface also works with models hosted in the company); bill reading uses a service in the EU region. Under the AI Act it most likely falls under limited risk: the assistant states that it is automated and that nothing runs without confirmation, and it never evaluates people.

Who they are: a company in the province of Padua, Italy, that treats waste, hazardous and non-hazardous. Trucks arrive with big bags, drums, IBC totes and pallets; then sorting, unpacking, shredding and baling.

The problem: for the waste register and RENTRI (the Italian national waste tracking register) they use the industry software that almost every plant has. It is old, some functions are badly made and support is practically non-existent: they told me a new function can take up to two years. And with waste, a mistake can end in a criminal complaint, even just for a missing label on a big bag.

What I built: a custom system for all internal waste movements, a kind of warehouse: deliveries are read from the industry software, every package has its own QR label, and you know where it is, what processing it has been through and who did it. It connects to the industry software, which still keeps the register and RENTRI.

How it went: it is in production, and the client told me it turned out to be a good piece of work.

Why it matters: the EU Battery Regulation will require the full history of every pack. Here it already exists: cell cartons, sorting measurements, tests, photos, defects and a digital passport in PDF, on a line producing 7,500 packs a year.

Before: functional tests, end-of-line tests and workstation photos in separate systems; defects had no history and every reject was explained by word of mouth.

What was built: a custom MES with traceability by serial number, structured defect events, a digital pack passport, statistics by workstation and by step, reading from and writing to the existing management system.

What changed: every pack has its passport and every reject a cause by workstation. The project is the basis of the European BattWISE proposal on adaptive work instructions (scored 12.3 out of 15) and of the visual quality study on 23,194 real weld photos: a study and a proposal, not delivered systems.

Data: runs on the company's server, networked with the workstations: photos and measurements never leave the plant.

Why it matters: in timber, margin is lost between the yard, the kiln and the saw. Here every load becomes a lot with its own code and every process records input, output and waste: you see where every pallet is and how much material is lost at each step.

Before: bill of materials, production orders and batches per job on separate sheets, with an administrative system nobody on the floor opened.

What is being built: bill of materials, production orders and batch traceability per job, with the same pattern (intake, batch, processing, output, label) already in production in the two cases above.

Status: in development, with a working demo every two weeks. Other systems in development, not described here yet: metalworking, industrial mechanics, solar energy.

Data: on the company's server, the client's code and data.

Why it matters: when an urgent customer calls, the delivery-date answer takes a minute, not a day: you add the hypothetical order and immediately see when it ships and which jobs slip, without touching the real plan. And the system tells you in advance how many orders to expect next month.

The problem: on the shop floor the plan is redone by hand for every new order, and nobody can say in advance whether a delivery date holds with the machines and shifts available.

What it does: production orders with phases, bill of materials and workstations; the plan respects shifts, maintenance stops, setup times and dependencies between phases, and shows on a Gantt chart by day, week or month. The what-if simulation adds a hypothetical order and immediately shows its effect on the whole schedule, without touching the real plan. Demand forecasting uses the order history (Holt exponential smoothing) and OEE measures availability, performance and quality for each workstation.

Where it started: built for a sheet-metal factory, then adapted to a metalworking company. It connects to the shop-floor machines. It is also the basis of the ZenithCore proposal to the European AID4SME call. The video uses test data.

Data and AI: runs on the company's server. Forecasting, OEE and simulation are statistics computed in house: no production data goes to external services, and a person always decides the plan.

Why it matters: with several stores the problem is not selling, it is knowing what is where. Here every garment that moves is a row in a single ledger: head office sees each store's stock and sales in real time, and each store sees only what it needs.

When: 2012, my first management system, for a fashion store chain based in Castelfranco Veneto, with a head office and affiliate stores. It stayed in use until the chain was sold.

Before: head office and stores worked on separate spreadsheets; stock, sales and documents were rebuilt by hand.

What was built: one system for head office and stores. Every movement of goods between stores and to customers is a row in a single ledger, which gives each store's stock in real time. Point of sale with correct VAT, invoices, delivery notes, credit notes and returns numbered per store and per year, price lists per store, loyalty cards with points and coupons, reorder suggestions from sales velocity, profit and loss per store and per day, sales by group and colour, internal chat between stores. Since 2015 it has also had a voice assistant: you talk to the computer to calculate VAT and discounts, open screens and take notes.

What changed: head office and stores look at the same numbers, documents come out of the system and no store sees head-office costs.

Data: a system installed for the chain, the client's code and data.

Why it matters: in a company with several sites an attack rarely shows up in a single alert: it shows up in behaviour. Here the system learns what normal looks like for each computer, by hour of day and day of week, flags what falls outside it, and an AI assistant suggests where to start the investigation.

My role: development team leader at a cybersecurity startup.

Before: at an industrial company with several sites, security alerts came from different tools with no single view, and every investigation was done by hand.

What was built: a single dashboard with devices on a map, the status of each one, alerts, process history, analysis of suspicious executables, IP checks against threat intelligence sources and automation rules that respond on their own to repeated events. An AI assistant reads the context of the alert and suggests what to check and how to act.

What changed: alerts sit in one place, background noise is filtered by the rules and whoever investigates starts from a suggestion.

Data and AI: data collection and learning of normal behaviour run on the company's servers; the AI assistant suggests, the investigator decides the actions.

Why it matters: sales are lost less to calls not made than to answers sent too late. Here every morning the CRM says whom to answer first, what to write and why, with advice taken from cited studies; the AI drafts the message in the customer's own words and a person checks it and sends it.

Before: a services company managed contacts and appointments on separate spreadsheets and calendars, with double-booked appointments, lost contacts and no measure of how campaigns were doing.

What was built: a CRM with a single contact register and appointment booking straight on the sales team's calendars, with no overlaps. After every call the operator records the contact's interest and notes, which stay for the later stages of the sale.

The AI inside: every morning the "Today" page lines up what to do, starting with whoever is waiting for an answer, with advice taken from cited studies and from the contact's own data. The AI drafts the message starting from the customer's own words, but a person sends it. An agent answers questions on the CRM data and proposes changes that apply only after confirmation. First-contact emails start from what the contact publishes on their website, with GDPR checks before sending, and incoming replies are classified automatically. The pipeline shows each deal's value weighted by its probability.

What changed: no more double bookings, each contact's history is in one place and every campaign has its own numbers.

Data and AI: the CRM installs on the company's server. The AI model connects to a provider chosen by the client (the interface also works with models hosted in the company), no message is sent without a person's confirmation and the agent's proposed changes apply only after approval.

Why it matters: the best tester's experience becomes guidance for every shift, and the photos the shop floor already takes become a quality check. The project was scored 12.3 out of 15 by the experts of the European AID4SME call.

The problem: on a battery pack line automated testing finds the defect, but deciding what to do depends on the experience of whoever is on shift, and that experience is lost between shifts.

What it does: every failed test opens a defect case by itself; the operator records checks, action and outcome, and sees which actions solved similar cases in the same phase. Work instructions show more or fewer steps depending on the measured experience of whoever uses them, and defect, decision and outcome go into the pack's digital passport. For the welds, 23,204 real photos from the line were analysed: a classifier built on DINOv2 sorts the photos into five classes with 99.4% accuracy on packs it had never seen. The full chain (quality filters, joint locator, anomaly models, verdict with heat map and confirmation by a person) is designed.

Evaluation: submitted to the European AID4SME call (Open Call 2, challenge C3.3). The experts scored it 12.3 out of 15, with the threshold at 9: implementation 5/5, novelty 4/5, impact 4/5, team 3.5/5. The video uses test data. Read the experts' evaluation (PDF, 2 pages)

Data and AI: everything runs in the company: photos never leave. The system flags, a person decides; it never evaluates workers (AI Act and Article 4 of the Italian Workers' Statute) and every decision stays in the log. The experts recognised the privacy-by-design approach and adherence to AI Act principles.

Why it matters: before putting AI into a company you need to know what the law requires, not afterwards. In ten minutes the tool states your role (provider or deployer) and risk level under the AI Act, the obligations that actually apply, each with its article and the date it applies from, and which paperwork you need: impact assessment, privacy notice, training register.

What is behind it: a registry of European AI model providers with the facts that decide a GDPR assessment: where data is processed, who the contracting party is, whether the provider trains on customer data, exposure to the US CLOUD Act. Twenty European providers and six non-EU providers with an EU region, each with the evidence level of its source.

Data: it runs entirely in the browser, the answers never leave the computer. It is not legal advice: it says what is needed and why; the paperwork is signed by whoever takes responsibility for it.