AI and data
LLM applications and AI agents.
We build assistants that answer from your documents with citations, customer chat, and agents that act with human approval.
Large language models (LLMs) are AI models that can read and write text. An AI agent connects such a model to tools, company data and business rules so it can carry out a specific task step by step.
01 / Definition
What are LLM applications and AI agents?
At Meybrain Technologies we build assistants that answer from your documents with citations, customer-facing chat that hands off cleanly to people, and workflow agents that act only with human approval.
A private assistant differs from a general chatbot in three ways: it knows your current documents, it respects who may see what, and it logs everything it does. Good agents do not try to do everything. They succeed at a narrow, measurable task with clear limits and a path to a human.
02 / Typical uses
Typical uses.
- 01
Internal knowledge assistant
Answers from procedures, technical documentation or regulations, citing the source.
- 02
First-line customer answers
Handles common questions and passes the rest to the right team with a summary.
- 03
Drafting
Proposals, reports or replies drafted from your templates and data, with a person making the final call.
- 04
Triage
Incoming emails and tickets classified and routed.
- 05
Tool-using agents
Checking an order status or creating a record within set permissions, with an approval step.
03 / Audience
Who it is for.
- Support, sales and HR teams answering the same questions every day.
- Organisations working with large document archives or fast-changing regulation.
- Software companies and start-ups adding AI features to their product.
- Companies that tried AI but could not trust the results enough to roll it out.
04 / Deliverables
What we deliver.
- 01
Use case and risk review
Which task suits AI, what an error costs and where a human must approve.
- 02
Knowledge preparation
Documents gathered, cleaned, chunked and made searchable.
- 03
Retrieval-augmented generation (RAG)
Answers grounded in your documents, with citations.
- 04
Tools and permissions
Which systems the agent can reach, and with what rights.
- 05
Evaluation set
Real questions with expected answers, re-run after every change.
- 06
Interface
A website widget, an internal panel or chat inside a mobile app.
- 07
Logging and cost tracking
Conversations, sources used and running cost.
- 08
Data-flow review
Against KVKK (Türkiye's data protection law) or GDPR where relevant, plus usage guidelines.
05 / Process
How the project runs.
- 01
One valuable task first
Not an assistant that "knows everything".
- 02
Evaluation set
Written before code, so success is defined up front.
- 03
Prototype
On your real documents, in users' hands early.
- 04
User testing
With the people who do the work.
- 05
Hardening
Permissions, approval steps, logging and failure handling.
- 06
Launch and monitoring
Quality and cost tracked in production; the evaluation set re-run on every update.
06 / Technology
Technology choices.
Choosing a model balances quality, cost and data location:
- 01
Commercial model APIs
(OpenAI, Anthropic Claude, Google Gemini and similar) usually offer high language quality and a quick start. Data is processed on the provider's infrastructure under contract terms that vary.
- 02
Open-weight models
(from the Llama, Mistral or Qwen families, for example) can run on your servers or in a cloud region you choose. You gain data control and take on more infrastructure work.
For document search we use vector search; pgvector on PostgreSQL is enough for most projects. For connecting agents to other systems we work with tool calling and current standards such as the Model Context Protocol (MCP). We avoid locking the application to a single provider, so the model can be swapped later.
07 / Example
Examples from our work.
Kentsel Dönüşüm AI is our own assistant for property owners going through urban renewal in Türkiye. It works alongside legal summaries, calculators and province guides, and is live on the web and the App Store.
On BeinStory, readers talk with book characters and authors through AI. We brought books, characters and authors into one content model and connected the conversation to it; we run the product ourselves on the web, iOS and Android. Both are in our portfolio.
In our portfolio
08 / Pricing
Pricing model.
AI projects carry two costs: building and running. Running cost is the model provider's per-token fee, or the infrastructure cost of a model on your own servers. During the prototype we estimate it from real usage samples and share the figure in writing.
Build cost depends on the number and format of knowledge sources, the systems the agent connects to, languages, evaluation depth and interface channels.
After a short call to settle the scope, we send a written, itemised quote for the build.
09 / Questions
Frequently asked questions.
- 01
What if the assistant gives a wrong answer?
Language models are not error-free and the risk cannot be taken to zero. We reduce it by grounding answers in your documents, showing sources, letting the assistant say "I don't know" and adding human approval for critical actions. The evaluation set checks that quality has not slipped after each change.
- 02
Will our data be used to train the model?
That depends on the provider's contract terms. We compare enterprise API options and self-hosted open models side by side and choose with you based on how sensitive your data is.
- 03
We have no AI specialists. Can we still do this?
Yes. We handle the technical side. We need the people who do the work to help build the evaluation set and test. We can also train your team on how to use the assistant and where its limits are.
- 04
Why not just use a general chatbot?
A private assistant answers from your current documents, respects user permissions, can connect to internal systems and keeps a record of every answer. General tools do not provide those controls in your environment.
- 05
Can one assistant serve several languages?
Yes. Modern models handle many languages, and documents in one language can answer questions asked in another. We include each language your users need in the evaluation set, so quality is measured per language rather than assumed.
- 06
How do you keep running costs predictable?
We measure cost per question on real samples during the prototype and project a monthly figure from expected volume. Caching, smaller models and tighter context keep it down.