Question list and document inventory
With your team we collect fifty to a hundred questions people really ask and mark which document answers each one. The list sets the scope and later serves as the yardstick for quality.
Service · Artificial intelligence
Cloud infrastructure for AI
Take a 25-person architecture and engineering practice in İzmir. Specifications, notes on building rules, soil survey summaries, old project reports and progress payment files have piled up for years in folders on the office server. A junior engineer wants to know which load assumption the team used last time on a comparable steel roof, and the one colleague who remembers is on site. The files belong to client projects, so feeding them to a public chatbot is not an option. What works is a cloud environment the practice owns. The model runs in the practice's account, the documents never leave it, and each answer shows the file and page behind it. We build such environments at European providers like Hetzner, OVHcloud and IONOS, or in the EU regions of Microsoft Azure, AWS and Google Cloud. Computing capacity is rented rather than purchased, and everything is configured remotely. Because processing data outside Turkey is a transfer abroad under KVKK, we work through those conditions with your legal adviser at the very start.
Your own account
model and files stay yours
Cited sources
file and page under every answer
Permissions
taken from SharePoint or the file server
Spending cap
alerts and automatic stop
What the work covers in practice
We design backwards from the questions. No server gets chosen until we know which questions will be asked, by whom, and which documents hold the answers.
Choosing provider and region
European providers and the EU regions of the large clouds are compared on price, GPU availability, managed model services and KVKK transfer conditions. You then choose on figures between a managed model service, an open model on a rented GPU, or a mix of the two.
Document pipeline
Text recognition for scanned PDFs, making tables readable, splitting files into meaningful sections, and tags for project, client and date. Where a specification exists in several revisions, the valid one is marked and older ones are removed from search or shown clearly as outdated.
Identity and permissions
Anyone barred from a folder in SharePoint or on the file server gets no answers drawn from that folder. Sign-in goes through Entra ID or whatever identity service you run.
Answer interface with sources
A chat window inside Teams or an internal web page. Under each answer sit the file name, page and document date, and one click opens the original.
Security
No internet-facing endpoints, encrypted storage, keys held in the key vault of your own account, and logs of who asked what and when, with a defined retention period.
Cost monitoring and KVKK paperwork
Budget caps, alerts, automatic shutdown after hours and a monthly usage report, along with a technical description of the data flow for your personal data inventory and your transfer assessment.
How we approach the job, from first call to handover
A first environment covering part of your files is usually running within a few weeks. Further growth follows data volume and what users tell us.
01
Questions and options
The question list, the document inventory and two or three architecture options with their costs.
02
Pilot
Files from one department or project group are indexed and scored against the question list.
03
User testing
A handful of staff use it in real work for two weeks; wrong or incomplete answers are collected and the pipeline adjusted.
04
Operation or handover
Ongoing care by us, or a handover to your IT team with the infrastructure documented as code.
Outdated versions cause more wrong answers than weak models do. When an old and a current revision of the same specification share an index, the model may quote either one and the reader will not spot which. That is a filing issue, not a model issue. We therefore tag every source with validity details and print the document date beneath each answer. The better organised your files, the less you need to spend on an expensive model.
Frequently asked questions
That falls under the KVKK rules on transfers abroad and needs careful treatment wherever documents contain personal data. Your legal adviser decides which transfer route fits; we document precisely where data goes, how it is encrypted and who can access it. Beginning with technical files that contain no personal data is another way in.
Usually not. When a question comes in, the system retrieves the relevant passages from the index and the model bases its answer on them. That is cheaper and needs no retraining when files change. Training only becomes a topic for very specialised language or formats.
Text, title blocks and notes on sheets exported as PDF can be made searchable. Interpreting the geometry of a drawing is a separate, costlier task, and we pin down expectations with real examples during the pilot.
Yes. The infrastructure is defined as code, and the pipeline and index are kept independent of the provider. A new model or region is tried against your question list and adopted only if it is provably better or cheaper.
Set-up is billed as a project or at €55 per hour plus VAT. For running costs you get a monthly estimate under several usage scenarios before starting, and cloud fees are paid by you directly to the provider. No hardware purchase is ever needed.
Neighbouring services that often go with this one
Service
AI adoption
Hand one business process to AI against a measurable target, from email inboxes to your TMS, from Outlook to Logo
Service
AI strategy
An AI plan with clear decision rights, a budget and written rules, aligned with KVKK and with what your customers expect
Service
AI training for teams
Live online seminars, department workshops, a shared prompt library and recorded modules for new starters
Plan an AI environment that belongs to you
Let us know which files should become searchable, how many people will use it and what you expect regarding where data lives. We will propose suitable options.
Availability
Weekdays 09:00-18:00 Turkey time (GMT+3); an answer follows by the next working day
Weekdays 09:00-18:00 Turkey time (GMT+3); an answer follows by the next working day
Calls
By video, over Microsoft Teams or Google Meet
By video, over Microsoft Teams or Google Meet
We have your enquiry
A reply will reach you by the next working day at the latest. If your message says work has come to a halt, it goes to the top of the pile.
- Filling the gaps. If we need more information to judge the job, we send questions by email or propose a quick Teams or Google Meet call.
- A written quote. It lists the scope, a euro price excluding VAT and a realistic start date. No hidden clauses, no items that appear later.
- Your decision. The quote arrives by email. Take whatever time you need, raise questions on any line, and decide when you are ready.