AI Development Services

We benchmark models. We do not chase the logo.

Building an AI diabetes coach, our engineers did not reach for the default. We scored several large language models on answer quality, speed, reliability, developer experience and cost, then picked the one that won on the metric that mattered for that product. It happened to be OpenAI. On a different job it has been Claude, and on another Gemini.

Choosing a model is an engineering decision with a cost attached, not a branding exercise.

Cost governance is designed in, not discovered on the invoice.

We own and operate Echo Journal, an AI voice journaling app on the App Store and Google Play. We pay its inference bill every month, which changes how you design. It runs a tiered pipeline. A cheaper model does the first pass cleanup, and the newest model is held back for the smaller, high value analysis where the quality difference is visible to the user.

If you just use the best AI for every part of transcription and summarizing and all of that, you are going to bankrupt yourself.

Dan Haiem, CEO, on building Echo Journal

The same discipline shows up in client work. On a document parsing feature we chose a lightweight model because the task was schema constrained and paying frontier prices per upload was not justified. We checked the file size cap twice, because a client reported size is easy to spoof, and rate limited parses per user per day. A 40MB PDF billed against your AI quota is a bug, not a feature.

Model-agnostic by architecture, so a provider change is not a rebuild.

On a K-12 education product we built a pipeline that assigns each job to the model best suited to it. One model generates the written content and rewrites it across a grade-level gradient. A separate, cheaper vision model scores how well candidate images match that content and filters anything that is not age appropriate.

The obvious alternative was licensing an existing dictionary API. We rejected it for two reasons. The phrasing was not pitched right for the student actually reading at that level, and the text is frozen, so you cannot regenerate a definition when classroom feedback shows it is not landing. The accepted trade is a higher per-request cost than a static lookup, in exchange for content that can be retuned per word and per grade on demand.

Cost discipline means matching spend to value, not minimizing spend.

We draw the decision boundaries before we write the prompts.

In the diabetes coach, we decided early what the AI was allowed to recommend and what had to stay deterministic and rule based. Health guidance is the obvious case, but the principle holds anywhere the wrong answer has a consequence. It also keeps the product usable, because a tight boundary produces quick, clear guidance instead of a chatty essay when someone is standing at a restaurant table deciding what to order.

That product went through App Store health review twice. We added the exact disclaimer language the reviewers asked for, were rejected again, then went back with our PM and lead engineer pointing the reviewer to the precise areas we had changed. Approved a couple of days later. Shipping AI into a regulated category is a process, and we have run it.

We cut the AI feature we were most excited about.

Echo Journal was meant to launch with a conversation mode, an open back and forth with the journal itself. It was the most compelling part of the vision. We held it out of version one, because it was also where the risk lived and where the bugs were hardest to reproduce consistently.

AI and wellness apps need extra care around trust and reliability, because a broken flow feels more personal to the user than it would in a standard utility app.

Any agency will tell you how fast they ship AI. The more useful question is what they refused to ship, and why.