Fast Web

Fast Web LabsWhat we test in hoursbefore selling it in weeks.

Experiments and client proposals: things we build fast to find out whether an idea holds up, how long it takes to stand up, and what makes somebody share it. If a client buys the proposal, it moves to Projects. Some are still running, some are archived, and some never reached production.

The ones that went badly are published too, and any figure that does not come from a client in production says it was measured in testing. A lab that only shows wins is a portfolio with another name.

2026

Build time
3 days
Status
Prototype
When
October 2026

Stack

Next.js · three.js · React Three Fiber · model-viewer

LAB / 008

3D configurator for manufacturers

A retail display and a branded terrace that the brand configures in 3D, switches between campaigns and places at real size with a phone. It comes with a plan, a timeline and a quote.

Open it
What we were testing
Does a display manufacturer close sooner if their client sees their brand on the product, at real size, before signing?
What happened
A demo with an invented brand: modular display and terrace, campaign switch, augmented reality without installing anything, and a proposal with a plan, a schedule and a sample quote. Just published: no results with manufacturers yet.
What we learned
The hard part was not the 3D: it was making the model, the plan, the schedule and the quote come from the same numbers. When they come from different places they do not add up, and a manufacturer spots it at once.
Build time
1 day
Status
Prototype
When
October 2026

Stack

HTML · JavaScript · Vercel

LAB / 007

Shelf plan for F&F

A proposal for F&F, who represent Finca Ferrer in Argentina: a kit to get into new supermarket chains and a page that tells shoppers which supermarket stocks each wine.

Open it
What we were testing
Does a distributor say yes sooner if it sees the proposal working, instead of reading a plan in a PDF with a quote?
What happened
Demo built on the real data from their sales deck: what the shopper sees, the kit for each chain, the weekly report and how it gets paid. Waiting on F&F’s answer.
What we learned
A plan in a PDF does not land: it needs something you can touch. And with no guarantees you do not charge much: the proposal became a small monthly fee, cancel any time.
Build time
2 days
Status
Prototype
When
September 2026

Stack

React · SVG · Vercel

LAB / 006

Shared capacity network

A system for health insurers to share capacity: an overloaded ER refers patients to a center with room, even if it belongs to another insurer, and settlement between them happens on its own.

Open it
What we were testing
Can the idle capacity of a health system in crisis (beds, appointments, maternity wards) become revenue without building anything new?
What happened
The prototype is complete: an occupancy map, four flows, screens for members, centers, treasury and leadership, and a revenue-split calculator. We showed it to people in the sector in Argentina. It did not move forward: nobody could pay for it, and no insurer wants to send its members to another.
What we learned
The blocker was not technical, it was about power: referring a member means losing them. The version that fits is the reverse: not “share your patients” but “fill your empty capacity”.
Build time
4 hours
Status
Running
When
August 2026

Stack

Next.js · Supabase · Vercel

LAB / 001

Farmear Aura

An aura test, a voted ranking and one-on-one battles, built while the trend was still hot.

Open it
What we were testing
If a mechanic is simple enough and lands in time, is somebody else’s trend enough to carry it?
What happened
Zero to production in an afternoon: domain, database, voting, battles with their own URL and a finished look. No viral traction.
What we learned
The speed is real and demonstrable. Somebody else’s trend is not distribution: with nobody to show it on day one, a good mechanic never starts.
Build time
2 days
Status
Archived
When
January – September 2026

Stack

Next.js · OpenAI · Supabase

LAB / 005

AI Architect

A strategy generator: you gave your trade, your city and your biggest problem, and in a minute it returned the infrastructure you needed.

What we were testing
Is giving away an automated audit enough for a business that does not know us to hand over its details and start a conversation?
What happened
It sat on the home page’s primary button for eight months. Nobody used it: not one report requested from cold traffic.
What we learned
We were asking before proving. Nobody tells you their biggest revenue problem ten seconds after landing on a site they do not know. And it was the one piece of the site that looked like every other AI agency. The same mechanic in reverse (a sum with your own number, no sign-up, nothing given in exchange) is what now lives at /tu-canal.

2024 – 2025

Before Fast Web. They worked end to end and none of them reached production.

Build time
4 weeks
Status
Prototype
When
2024 – 2025

Stack

React · Google Apps Script · Drive API · Tailwind

LAB / 003

Field work reports

An app for a technician to close the report on site: photos, customer signature and a PDF filed before leaving.

Open it
What we were testing
Can paper be removed from a field operation without asking the technician to learn anything new?
What happened
Working end to end: ID login, location lookup, signature and media capture, and a PDF generated and archived on its own. Never reached production.
What we learned
Adoption is not decided by the app: it is decided by how many taps it takes one-handed, wearing gloves.
Build time
Weeks
Status
Prototype
When
2024 – 2025

Stack

n8n · Gemini · Qdrant · Redis · Baserow · WhatsApp

LAB / 002

Catalogue agent over WhatsApp

An agent answering catalogue questions over WhatsApp by searching the real products, not what the model imagines.

What we were testing
Can a catalogue business handle its enquiries without the agent inventing products or prices?
What happened
In testing it resolved 85% of enquiries with sub-3-second answers, escalating to a human when unsure. Never reached production.
What we learned
Retrieval over the real catalogue is what separates a useful assistant from one that makes things up. The bottleneck was not technical: the project lost its owner.
Build time
Weeks
Status
Prototype
When
2024 – 2025

Stack

n8n · Baserow · OCR · OpenAI

LAB / 004

Document intelligence

Documents dropped into a folder come out as database rows, with the database itself as the only interface.

What we were testing
Can a small business stop retyping data from documents without buying an ERP or changing how it works?
What happened
In testing, extraction under 5 minutes per batch and 98% accuracy on the defined fields. Never reached production.
What we learned
The value is not in extracting: it is in going back and pulling a new field out of everything already processed without redoing the work.

Fast Web Labs

Got an idea nobody has built yet?

What we do in the lab is exactly what we do for a client, minus the invoice: understand the problem, stand it up fast, and look at the numbers without dressing them up.

Message us on WhatsAppWhat you pay