Skip to content
Résumé
FREN

N° 03 / 08 — SaaS · Voice AI

DeviJà

The quote writes itself while you talk: tradespeople dictate, DeviJà prices, sends and invoices.

Role
Design and development
Year
2026
Status
LiveProfessional
Stack
Next.js · TypeScript · PostgreSQL · Prisma · OpenAI · Stripe
← Back03 / 08
A dictation in DeviJà, and the priced quote it produces
Fig. 1What the tradesperson says, and the priced quote it produces.
DeviJà's home page on a phone: “Say it. Your quote is already done.”
Fig. 2On a phone, where quotes get dictated.
DeviJà's features: online signature, invoices, deposits and progress payments, reminders
Fig. 3Online signature, deposits and progress invoices, reminders: everything starts from the quote.

01 / 07

The problem

A tradesperson spends their evenings redoing quotes and invoices, when everything is already in their head: the job, their rates, their habits.

And electronic invoicing between businesses is becoming mandatory, for people for whom this is not the job.

02 / 07

What I did

The tradesperson says what needs doing. DeviJà identifies the client and items, asks for what is missing, calculates the quote, generates the PDF and sends it with an online signature link. Invoicing then takes one gesture: full, deposit, progress payments, balance.

The assistant learns their habits (“1 day = 9 h”) and stops asking questions it already has answers to.

03 / 07

How it's built

The principle: AI understands, the software calculates and decides. No total, VAT rate or document number comes from the AI.

It is a Next.js PWA with business logic kept apart from the interface, on PostgreSQL and Prisma. OpenAI handles transcription and structured answers; PDFs and Factur-X are generated on the server.

  1. Dictation
  2. Understanding (OpenAI)
  3. Calculation engine
  4. PDF, signature, invoice

04 / 07

Challenges

  1. 01

    Guardrails on the AI

    The AI never provides a total, a VAT rate, a company ID or an invoice number, and never issues anything on its own. Each of its answers is structured and validated (Zod) before the calculation engine takes over.

  2. 02

    Compliant e-invoicing

    Invoices are generated on the server in the Factur-X format (PDF and XML), deposits, progress invoices and credit notes included, and checked with the Mustang validator.

  3. 03

    An iPhone app, from the same code

    The PWA becomes an iOS app with Capacitor (files, sharing, push notifications), built and published by Codemagic. The website is containerised with Docker and deployed by GitLab CI.

  4. 04

    Getting known

    SEO is part of the product: a page per trade, guides, JSON-LD structured data and generated Open Graph images. Even the video ads are written in code, with Remotion.

05 / 07

Results

DeviJà is live: 20 free quotes and invoices, then €14 a month or €140 a year, with no commitment.

An account is ready in 3 minutes, and clients and rates can be imported from Excel or CSV.

06 / 07

What I learned

On a legal document, you do not hand the arithmetic to a model. Strictly separating what the AI understands from what the code decides made the product reliable, and far easier to test.

07 / 07

Tech stack

Front end
Next.jsTypeScriptPWAZodTailwind CSS
3D, motion and real time
Remotion (video as code)
Back end and APIs
StripeResend (email)Push notifications
Data
PostgreSQLPrisma
Mobile
CapacitorCodemagic
AI
OpenAI
DevOps, cloud and testing
DockerGitLab CI/CDGitVitest
Product, SEO and methods
Factur-X (e-invoicing)SEO

Next project · 04 / 08

JABS Software

Next project · 04 / 08

JABS Software

Write to me

contact@victor-delorme.com

Available for roles, projects and freelance missions