We build a custom AI assistant into your website, web app, mobile app or messenger. It runs on the OpenAI, Claude or Gemini API or an open-source model, not a no-code bot platform. We connect it to your data and set your rules in prompts and checks. It hands the conversation to your team when a rule says so or the model gives no answer.
We build assistants for three jobs: customer support, lead qualification and in-app guidance. We also build AI agents as separate projects.
Answers general and product questions from your own data, so your support team spends its time on cases that need a person. When your rules say so or the model cannot answer, the question goes to your team with the whole conversation.
Asks website and app visitors the questions your sales team would ask, collects contact details and recommends products. The qualified lead goes into your CRM, so sales sees the answers before the first call.
Helps a user finish a task inside your product in conversation, such as building a plan or a report instead of filling in a long form. The assistant runs the steps for the user in the conversation.
These are assistants and AI chat screens from published Ronas IT projects, including our own open-source app.
We built a cross-platform marketplace app in which an assistant on the OpenAI API helps buyers find plants that match their preferences. Buyers ask it about plant care from a plant's page or in a separate chat, and it suggests alternatives when a plant is not for sale.
We built and ship Open MobileUI, our own open-source React Native client for OpenWebUI. It is in the App Store and Google Play, and its code is public on GitHub under the GPL-3.0 license.
We designed an AI assistant chat for a neobank app concept. It lists the user's monthly subscriptions with the money spent, finds the nearest cash machines and takes voice messages that turn into text. The project paused after the design stage.
We designed a new version of the Mind Easy mental health app with the chatbot Olivia: users talk to it from the home screen and move around the app by voice. The app already existed, and development stayed with the client's team.
The assistant works where your users already are: on the web, in mobile apps and in messengers.
A chat inside your web app or on your site answers visitors on the page they are reading, so they get help without leaving it or writing an email. We add it to an existing product or build it into a new one with our web app development team.
The assistant lives inside your iOS and Android app, cross-platform or native, so users ask questions without leaving the app. Answers can arrive as chat messages with push notifications, so users see a reply even after they close the app. We build cross-platform apps with our React Native development team.
A bot in Telegram, WhatsApp, Slack or any other messenger with a bot API reaches customers in an app they already use, with nothing new to install.
The assistant answers from your own systems and writes its results to your CRM and support tools, so your team does not copy data by hand.
The assistant reads records from your product's database, so its answers use your current data rather than an outdated copy.
The assistant reads customer records from your CRM and support tools and writes results there, such as a qualified lead in the CRM.
The assistant connects through an API to other systems the job needs, so its answers can also use data from those systems.
Your application checks permissions before data reaches the model or an action runs, so the assistant uses only what that user may see and do.
The assistant hands a conversation to a person when your rules say so or when the model cannot produce an answer.
Your team member continues in the same conversation and sees everything the customer has written, so the customer does not repeat the question. The assistant can also add a short summary of the request, so the person who takes over knows what is needed before reading the history.
You set the rules, for example a topic the assistant should not answer, a customer who asks for a person, or a question outside the data it reads. The conversation goes to a support employee in your support tool or to a manager in the site chat.
We choose the large language model (LLM) for your budget and task and put guards around it.
We work with OpenAI, Anthropic Claude, Google Gemini and open-source models. We choose by your budget, the task's complexity, the input types such as images or video, and the features the assistant needs.
We run open-source models on servers you control, for example when an API costs too much at your volume. We fine-tune a model when prompts alone do not give the answers you need.
Prompts set the assistant's tone, topics and length limits and give a fixed reply to out-of-scope questions. This is a large part of the work. Prompts are versioned, and your team can edit them in the admin panel.
Code around the model allows only one pending answer per chat and sets a time limit for each request type. It treats a safety-filter stop or an empty answer as a failed call, so the app handles the failure instead of showing it as an answer.
We work out the model running costs from expected token counts and provider prices before we build, so you choose a model with an estimate of what it will cost to run.
A hosted model processes your data under its provider's terms, so those terms are part of the model choice. A self-hosted model keeps the data on your servers.
Six steps take the assistant from scope to support, and its answers are checked on your own examples before launch.
We agree on the job the assistant does, its channels, the systems it reads from and the rules for handing over to a person. You get an estimate of the cost and timeline.
We write the conversation flows, including replies for questions the assistant should not answer, and the prompts that set its tone and limits. The guards around the model are defined at this step too.
Our designers create the chat screens for your web app, mobile app or messenger flow so they match the rest of your product.
We agree with you on a test set of real questions and the answers you expect, then compare models and prompts on it. We fine-tune a model where prompts are not enough. We move to the next step when its answers pass your examples.
We connect the assistant to your databases, CRM and support tools, add the permission checks and cover the model calls with automated tests, including tests for provider failures.
We release the assistant in the agreed channels and keep improving its answers after launch, using the reports from your team. Support terms and response times are agreed in the support agreement.
We build the assistant on Laravel or Python and connect it to the model provider your task needs.
OpenAI, Claude, Gemini and open-source models, called directly or via OpenRouter, a gateway that lets us switch models without rewriting code.
Laravel and Python. In Laravel, the neuron-ai framework and queued model calls with timeouts, so a slow model does not stall the app.
Next.js for web AI features, with model calls made only on the server, so your API keys never reach the browser.
React Native for cross-platform apps, plus native iOS and Android apps, with replies delivered as chat messages and push notifications.
Where your backend is Laravel, our package ronasit/laravel-chat can run the conversation, so the assistant reuses chat code that already works.
The admin panel your product already has, Laravel Backpack or a custom one, where your team manages prompts.
We price an assistant inside your existing product as one AI feature. A new product built around AI is priced as an AI product from scratch.
Evgeny Leonov, CTO of Ronas IT, leads our AI work and has guided more than 120 apps from idea to release.
Before we write code for an assistant, we work out what each answer will cost on the models that fit the task. A client should see an estimated model bill before launch, not a surprise after it.
Then we check the answers on the client's own questions. An assistant that sounds right is easy to build. One that is right on the client's data is the real work.
These Clutch reviews come from integration and product projects, the work closest to building an assistant.
Numbers
