Web apps and AI products

A website tells someone about your business. An app does something for them.

We build the second thing: tools with a working brain behind the interface, shipped to real users rather than demoed in a deck.

What we build

Progressive web apps. Installable to a home screen, working offline, updating without an app store review. For most products this is a better answer than native — it ships in weeks instead of months and reaches everyone at once.

Internal tools and dashboards. The unglamorous software that decides whether your team's day is good. Usually the highest-ROI thing a small business can commission.

AI-backed products. Retrieval over your own content, structured extraction, assistants that are scoped tightly enough to actually be correct. We build the unglamorous parts — evaluation, guardrails, fallbacks — because an AI feature that is wrong 5% of the time in public is a liability, not a feature.

Content and publishing systems. Editorial workflows, membership and paywalls, newsletter infrastructure.

Shipped

• Cuyar — a bilingual movement app for adults over 65. Installable PWA, 29 illustrated movements, offline voice coaching, and a design built around the accessibility needs of the actual audience.
• Aguante — a bilingual fasting app whose entire thesis is that the popular claims in its category are not supported by the evidence. Research is a feature.
• Pour AI — a wine assistant you can ask a question or show a photo of your dinner, scoped tightly enough to stay trustworthy in public.
• News4You — local journalism with membership, on a stack a two-person team can actually run.

How we think about AI features

Most AI product work fails in the same place: the demo works, and then real users arrive with real inputs.

So we build in this order — decide what "correct" means and how we will measure it, constrain the model to a scope it can actually be right about, ship a fallback for when it is not, and only then make it feel good. Anything else is a prototype wearing a product's clothes.

We will also tell you when your problem does not need a model. A lot of "AI features" are a database query and a well-written rule set, and the version without the model is faster, cheaper, and correct every time.

PWA or native?

We build PWAs by default and say so when a project genuinely needs native — deep hardware access, background processing, or App Store distribution as a business requirement.

What it costs

Web app and AI projects start around $15,000. The range is wide because "app" covers a two-week internal tool and a nine-month product. We scope in phases so you get something usable early rather than a big-bang launch. See pricing.

Common questions

Can you take over an app someone else started? Often. We read the code before quoting, and we tell you honestly whether continuing or restarting is cheaper. Sometimes it is restarting.

Do you do the backend too? Yes. Typically Supabase or Postgres with serverless functions. We set up row-level security properly, which is the thing most rushed projects skip and later regret.

What about my users' data? Collected minimally, stored where you control it, documented in a privacy policy that matches what the code actually does. We have shipped consent flows and deletion paths for products in regulated-adjacent categories.

Will it work offline? If it should. Cuyar works fully offline because its users are often in basements and community centres with bad signal. That was a product decision, made first.

Can it be bilingual? Yes, and it is cheaper to decide that on day one than to retrofit it. See bilingual work.