Loading model
One company at a time. It reads your stock, your price lists, your CRM, and runs on your own machines. People included.
| PROCESS | WHERE | RATE |
|---|---|---|
| scintilla | MacBook Pro | 166 tok/s |
| imagineer | MacBook Pro | 1.4 s / image |
| eugeny | MacBook Pro | 27B in 7.6 GB |
| tomasi.feed | 5 showrooms | 850 pages |
| nims.leads | Meta · Google · TikTok | 3,275 / 10 days |
| pam.app | iOS · Android | 500K+ / downloads |
Tomasi Auto
Every car in stock gets a page in each of the five showrooms and its own ads, made from the car's real data and photographs. A car sells and its page and ads come down. A person approves every batch before it reaches Meta and Google.
170 cars · 850 pages · 2,979 form responses in June 2026 at €42 each
Read the case
Pam Panorama
For every flyer in every store, a model reads what's on offer and picks the recipes that use it: about a dozen out of a thousand, only real products and real recipes. Avocado at €0.99 in Milano Forze Armate? California roll.
500,000+ downloads on Google Play · 4.6 stars from 9,219 App Store ratings · 103 stores in the July 2026 calendar cohort
Read the case
Lavazza Group · NIMS
NIMS sells Lavazza coffee and water systems through more than 3,000 Personal Shoppers. We run the campaigns on Meta, Google and TikTok and deliver every request in the exact file the network has always used.
8 years · 5 product lines · 3,275 Meta form responses from 24 August 2026
Read the case
Viessmann
Vitoadvisor sends every customer request to a partner chosen by product skill and distance, then follows the quote and the site visit. Around it: partner portals, calculators, campaigns and films, some with sets and voices built by AI.
240 leads · 125 site inspections in the 2024 campaign · since January 2019
Read the case
Il Valutatore
One sentence, and the model finds the exact car in a catalogue of 308 makes, then prices it from twenty real comparable listings.
308 makes · 20 comparables behind each price
Read the case
Gateway · Eugeny
A private model on your own servers, or on the laptops your colleagues already carry. Everyone signs in with the company account, sees only what their role allows, and every action is logged.
runs on your servers or a laptop · permissions per person
Read the case
fornace.net
The internal app, the client portal, the dashboard nobody had time to build: describe it on fornace.net and it plans, codes, tests and publishes it while you watch.
live at fornace.net
Read the case
AIA · Wudy · Pam · Damigella Blu
We flew to Tenerife for Wudy. We filmed Damigella Blu's wine underwater. We made Pam's film about its fishing boats and the Valpolicella documentary Il Classico Speri.
[Play reel]
Read the case
Sunamp
Every lead tells us its trade, and that label goes back to Meta and Google, so the platforms learn which clicks become installers. Since May 2026, installers have been the largest group in the UK and in Italy.
summer 2025 · 130 leads, 48 installers in training, 19 days
Read the caseEfficient enough to run on your laptops: no server racks, no data-centre graphics cards. Our models work on the Macs you already have, even with the Wi‑Fi off: on a train, at a client's, in a plant with no signal.
fornace.net is fifteen years of making sites and apps, turned into a machine: you describe what you need, it builds it. People have built a playable voxel world on it, a sprint-training app and a property research tool. We built a wine shop for Bonazzi there.
Build something on fornace.net ↗
Everything below was built in Valpolicella, down to the code that runs it on a laptop chip.
| scintilla | language model with its own engine, runs on any Mac with an Apple chip | 166 tokens/s on a MacBook Pro |
|---|---|---|
| eugeny | a big model, compressed to fit on a laptop | 27B parameters in 7.6 GB · in training |
| imagineer | images and video on a Mac, offline | one image in 1.4 s |
| torchio | compresses models to 2 bits and keeps what they know | 2.2× more accurate than plain 2-bit rounding |
| forgia | where our models get their data, their training and their exams | every step on record |
| gateway | one door to every AI model, with budgets per person and automatic failover | open source |
| matchbox | identity, permissions and secrets, so nobody carries a password | 21 connectors |
| fornace.net | describe it, it gets built | live |
MacBook Pro M5 Max, 48 GB. Imagineer: 512 px image, 4 steps, warm. Scintilla: q8, short context.
Your colleagues sign in with the company account and each finds the AI their role allows: sales the CRM, accounts the ledger, the new hire an onboarding guide to ask anything. Every model behind one door, every action logged, no password in anyone's hands.
09:02 giulia sales crm.search ok09:03 marco new hire onboarding.ask ok09:05 agent:ads marketing meta.upload approved by luca09:06 marco new hire ledger.read denied
We write the messages, choose the frame and sign off what goes out. Next to the office there's our set, with a 50 m² kitchen, cinema cameras and drones.
Fornace was founded in 2011 by Francesco Frapporti and Luca Cisorio. It began with native iPhone apps for marketing. Today it's 20 people plus a circle of collaborators, in Valpolicella.
Book a call Send this to your IT
Keep it. It has never seen your stock, your price lists or your CRM. We build the part that has.
No. Scintilla, Eugeny and Imagineer are ours: we train them, compress them and wrote the engines they run on. When we use someone else's model, it goes through the same door as ours.
Not with your data. The model trained on your work, and the code around it, belong to you. Your training data and every answer the model gives stay in a chain of custody we sign with you. It's in the contract.
Only if you send it there. Our models run at your premises, in approved European data centres, or on one laptop with the Wi‑Fi off.
No. To start, a Mac from a shop is enough. If you'd rather keep it in a rack, we host at your premises or in approved European data centres, with the chain of custody in writing.
Neither is Tomasi. Five showrooms, 850 pages, and the stock keeps them current.
Code checks what code can check, and a person signs off what goes out. At NIMS, AI never produces a number: the code counts the leads.
Before we build, we write down which obligations fall on you and which on us. Generated content is labelled, every action is logged, permissions are set per person. When your lawyer arrives, the answers are already written.
Yes, and you take everything with you: models, data, code. Whoever comes next starts from there, not from zero.
Good. Here's the page for them.
Scope and fee are fixed in writing after the first call.
Usually Francesco or Luca.