Services · AI systems
We use the model that fits the job, and we build the system around it: the data, the tools, the limits and the people who approve. We build it, and we stay to run it.
AI systems
From one LLM call to a team of agents that runs a platform.
We connect your product to Claude, GPT, Gemini, Grok or an open model, through the direct API or through OpenRouter. One adapter talks to all the models, so a change of provider is a change of configuration.
How we do it : LLM integrationAn assistant that answers from your documents and your data, remembers the conversation and shows its sources. We add guardrails, usage limits and cost control.
How we do it : Chatbots on your own dataAgents research, plan and prepare changes on a schedule. Each result is a proposal with its reasons, and nothing happens until a person accepts it.
How we do it : Agents that propose, people who approveA team of always-on agents, each with one role and its own rules. One network develops, maintains, deploys and monitors a platform. Another turns a raw recording into a finished media channel.
How we do it : AI agent networksWhen the data cannot leave the building, the model comes to the data. We run open models and local embeddings on your servers, and we anonymize sensitive data.
How we do it : Private AI on your own hardwareModels that watch a live stream and act. They mark the events of a match, or they keep the play in the frame of a camera.
How we do it : Computer vision on live videoMethod
We do not start from one provider. We test a prompt or a model against an evaluation set before a change goes live.
An agent can research, plan, code and test. A person approves each task, each plan for a complex task and each production deploy.
You can see what each answer cost and why the model gave it.
The work does not stop at the release. We run in production the systems that we build.
Tools
Proof
Each project has a page that says what the product does and what we built.
Fintech SaaS · AI research agents
A portfolio and dividend tracker with AI research agents. The agents propose, and the investor approves each move.
Health AI assistant
A personal medical assistant that reads your records and answers daily health questions with your own history in mind.
Autonomous media channel · AI agents
A role-play channel that AI agents operate from end to end: they edit the episodes, cut the shorts, publish them and maintain the web site.
Document chatbot · On premise
A private chatbot that answers from the documents of a company, on the servers of the company.
Computer vision · On premise
A camera that follows the play by itself: AI detection and PID control, live and on premise.
Multiplayer browser RPG
A pixel-art online RPG. Belay Studio directs and designs the game, and AI agents write the code and the architecture with our technical direction.
Questions
Short answers, before you write to us.
Claude, GPT, Gemini, Grok and open models. We use the direct API of each provider (Anthropic, OpenAI, Google, xAI) or OpenRouter. Our services talk to the models through one adapter, so a change of provider is a change of configuration.
Each model call has a trace and a cost. We set usage limits and budgets, and we test a prompt or a model against an evaluation set before a change goes live.
You do. Each agent has one role and its own rules. In a software team, a person approves each task, each plan for a complex task and each production deploy. In vest101, each proposal of an agent waits in an inbox until the investor accepts it.
Tell us what you want to build or fix. An engineer answers.