LLM integration and AI assistants
AI that sees no more than your user
We add assistants, document extraction and agents with fixed limits to your product. The model works with the signed-in user's permissions and gets only the context it needs. Tests check how it behaves.
AI
What we build on large language models (LLMs)
01
An assistant inside your app
The assistant answers questions about your data and can create or update records in the app. Every change shows up as a preview first and runs only after the user confirms it.
Stack
- Vercel AI SDK
- Gemini
- Claude
- Next.js
- NestJS
02
Data from documents and images
We pull structured data out of PDF invoices and check it against a schema before it is saved. Images get tags from a fixed list, and tags the model is unsure about go to a person for approval.
Stack
- Gemini
- Vercel AI SDK
- Zod
03
Agents within fixed limits
The model writes no SQL of its own and has no direct database access. It calls only fixed tools on top of your API, and its output, such as a dashboard draft, has to pass a schema before the user sees it.
Stack
- Vercel AI SDK
- Zod
- GraphQL
- Claude
04
Streaming chat interfaces
The model's answer appears in the interface as it is written. While it streams, we lock the actions that would send a second request, and bulk changes show up as a clear preview.
Stack
- React
- Next.js
- Vercel AI SDK
- SSE
Process
How we keep AI within limits
Tools run as the user
Every tool call goes through your regular API with the signed-in user's token, so the same permissions apply as in the app. Tools the user may not use are never given to the model.
Context on demand
The prompt lists only tool names and short descriptions, and the model loads a full tool when it needs it. For one assistant this cut the fixed context per request by an estimated 60%.
Behaviour checked by tests
We cover the assistant's tools with tests and try the permissions in the browser as a user who has no access to the data. We set the threshold for automatic decisions from the precision measured on your own data.
When AI makes sense and when it does not
Makes sense
- Reading the same kind of document over and over, such as PDF invoices.
- Answering from data the user is allowed to see.
- Proposals that code, a schema or a person then checks.
Does not make sense
- A task a plain rule or filter already handles.
- An answer where a mistake cannot be checked.
- A decision that would bypass the user's permissions.
Experience
AI features we have built
We don't name clients. Each entry gives the type of product and our work on it.
01
Legal e‑billing, US
Full stack on Next.js, NestJS and PostgreSQL. We took over the AI assistant: its tools run as the signed-in user and load only when the model asks for them, which cut the fixed context per request by an estimated 60%. Also unit tests and CI for the web app.
Stack
- Next.js
- NestJS
- PostgreSQL
- Vercel AI SDK
- Gemini
02
Roofing company
Frontend architecture and guardrails for an MVP first built with AI. LLM image tagging and invoice extraction, built with one other developer.
Stack
- React
- Claude Code
- LLM
03
Legal AI and fintech
Frontend and streaming chat interface for an LLM tool for attorneys. The foundation of the commissions app in a platform of nine Next.js apps.
Stack
- Next.js
- React
- LLM
04
Grammar-checking SaaS
Most of the application, backend included: an API on Fastify, Prisma and PostgreSQL, a Crawlee crawler, and checks routed by language to LanguageTool or GPT-4o-mini.
Stack
- Fastify
- Prisma
- PostgreSQL
- Crawlee
Frequently asked questions
Can the AI see data the user cannot?
No. Every tool call goes through your regular API with the signed-in user's token, so the same permissions apply as in the app. Tools the user may not use are never given to the model.
How do you check that the AI behaves correctly?
We cover the assistant's tools with tests and try the permissions in the browser as a user who has no access to the data. We set the threshold for automatic decisions from the precision measured on your own data.
Technologies
- Models
- Gemini, Claude
- Model integration
- Vercel AI SDK, Zod, GraphQL
- Application
- React, Next.js, NestJS, SSE
We add AI to the applications we cover with web application development and to stores we cover with Shopify development.
Contact
Let's find where AI fits your product
Write to us and we will go through which steps a model can take for your users, and with what permissions.
What to put in the message
- What you want to build or change.
- The systems and data you already have.
- Constraints, such as a deadline and who will maintain it.