Applied AI
AI App Development
Elermond Studio builds AI applications that run in production: chat and voice agents, assistants that read email and calendars and take actions, retrieval over your documents, and AI features inside Flutter and web apps. We work with Claude, OpenAI and open models, and we treat cost, latency, evaluation and safety as part of the product, not an afterthought.
What we build
Customer-facing agents: chat and voice assistants on your site, in your app or on Telegram and WhatsApp, grounded in your knowledge base (UnicFlux).
Internal assistants: tools that unify email, calendar, CRM and tasks into one conversation with a confirmation step before any action (NeoCEO).
AI features in consumer apps: photo-to-nutrition estimation, personalised plans and coaching inside a Flutter app (EatingAI); an AI finance assistant that proposes edits you confirm or undo (MyStart).
How we keep it reliable
Every prompt and tool has an evaluation set. We measure answer quality, refusal rate and cost per task before and after each change.
Guardrails where money or data move: confirmation steps, allow-lists for tools, rate limits, audit logs. Prompt injection from documents and web pages is treated as a security boundary.
Model routing: cheaper models for classification and extraction, stronger models for reasoning. Caching and streaming keep latency and bills predictable.
Stack
Claude and OpenAI APIs, Vercel AI SDK, Node.js and TypeScript backends, PostgreSQL with pgvector, Supabase or Firebase, Flutter and React front ends. Deployed on Vercel, Railway, Fly or your cloud.
Built by us
Questions
Which models do you use?
Whichever fits the task and your data policy. Claude and GPT models for reasoning and agents, smaller or open models for classification. We can keep data in the EU or on your own infrastructure.
Can you add AI to an app we already have?
Yes. See the AI assistant integration page: we add an assistant or an AI feature to an existing Flutter or web app without a rewrite.
How do you control API costs?
Model routing, prompt caching, token budgets per user and per feature, and dashboards that show cost per task from day one.
Is our data used to train models?
No. We use API tiers with no-training terms and document which data leaves your systems and where it goes.