Scentena: an AI perfume formula generator that respects IFRA limits

Scentena is an AI perfume formula generator for hobbyists, students and indie perfumers, built and run by our co-founder Anis Setoodeh. It generates formulas by theme, product type and composition method, applies the IFRA Category 4 safety limits automatically, and turns a plain-English brief into a formula you can weigh out in grams. It is live at scentena.com. This case study covers the problem, what Anis built, the engineering behind it, and what it says about how ELORA DIGITAL approaches AI features.

Perfume formulation has rules. Each material has a maximum safe concentration under the IFRA standards (for fine fragrance, the limits in Category 4), a formula has to add up to exactly 100 percent, and the balance between top, heart and base notes decides how long a scent lasts and how far it projects. Hobbyists and students learn these rules the slow way, one failed batch at a time, and the spreadsheets they build to track them are fragile.

A general-purpose AI chat is tempting here, and risky. Ask it for a formula and it can produce confident percentages, name materials that do not exist and ignore the safety limits. The problem Anis set out to solve was not "generate a formula" but "generate a formula that cannot be wrong in the ways that matter".

• Formula generation by theme, product type and composition method, with longevity and projection optimisation.
• Safety built in. IFRA Category 4 caps are applied automatically to every formula, and the result is given in gram weights so it can be weighed out at the bench.
• Two classical methods. The Jean Carles method and the pyramid method, for people learning the craft the way perfumers teach it.
• The AI Lab. Write a brief in plain English, such as a light cologne for summer, and get a buildable formula back.
• An ingredient builder for custom materials and accords, on top of a database of about 95 materials.
• A saved library (up to 10 formulas on the Free plan) and CSV export, so a formula can leave the app for a spreadsheet or a supplier order.
• Thirteen learning articles that explain the methods and the rules.

Plans are Free at $0 and Pro at $9 a month, billed through Stripe, with Google and GitHub sign-in.

For a business reader: Scentena is a fast single-page web app with an API behind it, a database that can run small (SQLite) or scale (PostgreSQL), AI from Claude, and Stripe for billing, hosted on Vercel and Fly.io. For a technical reader: React 18 and Vite with build-time prerendering, Tailwind and shadcn/ui, an Express API in TypeScript, SQLite or Postgres, and Claude with a smaller model on the Free plan and a larger one on Pro.

The engineering decisions are where the product's honesty lives:

• The AI chooses, the engine decides. The model may only pick from a fixed list of real materials. A deterministic engine then enforces the IFRA limits and normalises the formula to 100 percent. The model cannot hallucinate a percentage because it never sets one.
• Credits are reserved, then refunded on failure. An AI request reserves the user's credit up front and gives it back if the call fails, with a 60-second timeout so a slow response cannot hang the app or quietly charge for nothing.
• Login lockout after five failed attempts, to slow down password guessing.
• A prerender bug found and fixed. Build-time prerendering makes the site fast and indexable, and in September 2026 it exposed a hydration bug for users with a saved dark theme. It was traced and fixed. Small, visible bugs like this are the reason a care plan exists.

AI with guardrails. The pattern in Scentena is the one we apply to the AI features we build for clients: the model proposes from a fixed set of real options, deterministic code checks the result against the rules, and a person sees the outcome before it matters. It is the same shape as Reckonly, where AI reads a receipt and a person confirms it before anything reaches the accounting software.

Costs and abuse considered before launch. Credit reservation, timeouts and lockouts are not glamorous, and they are the difference between a demo and a product people can pay for.

A product people can learn from. The thirteen articles and the two classical methods are there because a tool that teaches its users is worth more than one that only answers. We take the same view of documentation and handover for client projects.

Shipping over time. Scentena started in August 2025 and most of the work happened in June 2026. Products are rarely built in one go. What matters is that each release is complete, tested and safe.

Scentena is a product Anis built and runs himself. It belongs to him, not to ELORA DIGITAL, and it is not client work. No user or revenue figures are published for it, and none are quoted here. ELORA DIGITAL opened in October 2026, and client case studies will appear as projects launch and each client agrees in writing to share their story. In the meantime, if you have an idea for a product with AI in it, how much does a custom web app cost explains what a first version costs, and you can tell us about it in your own words.