▶ In-house AI assistants
Your knowledge problem is not always a search problem.
I build an AI assistant that works on your procedures, reports and past projects, answers with sources, and is mathematically optimized for speed and low cost.
The knowledge base I built for Sıfır Atık Vakfı (Zero Waste Foundation) is one example: the foundation can query its past projects, reports and relevant world developments in one place.
Atalay Deveci · Updated 2026-10-08
What this service actually is
An in-house AI assistant connects large language models (LLMs) to your company's own documents. An employee asks a question; the assistant finds the relevant documents, writes the answer based on them, and shows where it came from.
It is not pasting company documents into a public chat tool. The assistant works with your company's data and is mathematically optimized to answer fast and run at low cost.
The uncomfortable part
Not finding information usually looks like a search problem. This is also where most AI projects get stuck.
The documents are outdated or contradict each other.
If there are three versions of the same procedure, the assistant gives three different answers. AI doesn't organize scattered knowledge; it makes it quick to find.
There is an assistant, but nobody trusts it.
An assistant that shows no sources and doesn't admit when it's unsure gets abandoned after its first wrong answer. Trust comes from seeing where the answer came from.
The pilot never ends.
An AI trial with no success criteria defined up front can run for months without reaching a conclusion. Without clear scope and a measurable goal, a pilot stays an experiment.
I will tell you early which of these applies. If I see your problem is solved by a good document structure rather than an AI assistant, I will tell you that too.
▶ How I work
- Week 1
Use case and sources.
Who is the assistant for, and for which questions? Which documents are in scope, and which are current? How success will be measured is written down here.
- Weeks 2–3
First working version.
Documents are ingested, the assistant is set up and answers with sources. It is tested with real users' real questions.
- Weeks 4–5
Accuracy and trust.
Answers are measured regularly against a question set. Wrong or incomplete answers are fixed, and the assistant is made to say when it doesn't know.
- Final week
Handover.
How to add new documents, how to monitor the system and how to guide users is documented and handed to your team.
▶ When you should not hire me for this
If any of these is true for you, let's not start this engagement:
You have few documents and everyone knows where they are.
A well-organized folder structure and search may be enough. An assistant would be an unnecessary layer.
The knowledge is in people's heads, not in documents.
No assistant can find knowledge that was never written down. It has to be written down first.
Answers must be error-free.
Where errors are unacceptable, such as legal or medical decisions, an assistant can only support an expert, never replace one.
The goal is just to say "we have AI too".
An assistant built without a specific business question soon goes unused.
▶ Engagement shape
We start with a fixed-scope pilot of 4–6 weeks. The pilot covers one team and one document set, with how success will be measured written down up front.
At the end of the pilot you have an assistant real users are using, with measured accuracy. Whether to continue is decided on that result.
- 30-day exit
- If it isn't working, you can end the engagement with 30 days' notice.
- Invoiced from Türkiye
- I invoice through my company registered in Türkiye.
- Reply within 24 hours
- I reply to every message within 24 hours.
▶ Questions and answers
Which language model do you use?
The model is chosen for the use case, data sensitivity and your infrastructure. We make that decision together, with the reasoning on the table.
Does it work with Turkish documents?
Yes. Turkish and English documents can be used together; the assistant answers in the language the question was asked in.
What happens if the assistant gives a wrong answer?
Every answer comes with its source, so the user can check it. During the pilot, answers are measured against a question set and improved.
What happens when new documents are added?
How to add new and updated documents is part of the handover to your team.
▶ Fastest route
Start with one email.
The link below opens an email with the questions that make the first conversation useful already written in. No form, no account. I reply within 24 hours.
hi@klarvoanalytics.com
▶ Other services
- Corporate trainingsThe team can't work with data.
- Demand and sales forecastingNobody can predict what will sell.
- Customer analyticsLost customers are noticed too late.
- Reporting and data infrastructureReports are built by hand and take days.
- Custom dashboardsOff-the-shelf BI tools can't show what we need.
- Fractional data science leadNobody owns data.