AI solutions
AI assistants
A helper inside your product or your team chat: answers questions from your data, drafts replies, suggests the next step. It answers from what you gave it, not from what it invented.
- Reply to your brief
- within 24 hours
- First call
- 30 minutes, no commitment
- First working version
- in 3 weeks
Process
What happens, and when
The work is split into stages, and each one has a named result. You see it yourself rather than reading about it in a report.
Days 1–4
Answer boundaries
We collect the questions people actually ask you — from email, tickets, the team chat — and sort them into three piles: answerable from your documents, answerable nowhere, and questions the assistant must not touch. The third pile matters most: it sets where it stays quiet and fetches a person.
You get — The questions sorted and where it stays quiet
Week 2
Placement and permissions
We settle where the assistant lives — a panel inside the product, a bot in the team chat, a search box — and what it can see. Permissions are checked before the model, not after: nobody should read a document through the assistant that the system keeps from them.
You get — A chosen home and permissions wired in
Week 3
Answers with sources
Answers are built on top of your own documents, each carrying the passage it came from. We test on week one’s questions, and separately on the ones whose correct answer is “that is not in the documents”. Knowing when to stay quiet matters as much as answering.
You get — Answers citing the passage, tested on real questions
Before launch
Your own people first
It opens to your own people first — support, sales, whoever spots nonsense before a customer does. Every answer carries a “this did not help” control, and those questions arrive as a list rather than dissolving into a chat thread.
You get — The assistant with your team and its misses
Every month
Misses and sources
We look at what people ask and where the assistant misses. More often than not the fix is not the prompt but the source document: the answer genuinely is not in it, or it is written so that no person would call it correct either.
You get — A report on misses and edited source documents
What the work covers
Answer boundaries
what the assistant answers from your documents, what it hands to a person at once and what it never touches — the list comes out of your tickets and threads rather than off the top of anyone’s head
Building the knowledge base
an export from the wiki, shared drives and mail, split into passages, search by meaning, re-indexing when a source is edited — what leaves is the passage for a given question, never the archive
Answers with the source
every claim shows the document and the paragraph it came from, and with no suitable passage the answer is “that is not in the documents” instead of a plausible guess
Permissions first
wiring to your roles and groups, closed documents cut off inside the search itself, a check before an answer goes out, and a log of who asked what
Panel, bot or search box
built into a product screen, a bot in the team chat, or a search field — a different amount of work each, and a different amount the assistant can see about who is asking
Where it missed
a “this did not help” control next to each answer, a report on the questions left unanswered, and edits to the documents that make it get things wrong
What we need from you
- Access to the wiki, the shared drives and the databases the assistant answers from
- Someone allowed to rewrite an out-of-date document and settle two versions that disagree
- A decision on whether it answers your staff or your customers — the bar for when it must stay quiet differs
Projects in this area
Logistics
Vezira
A haulage company with a mixed fleet: the assistant reads a dispatcher’s request, gathers data from the CRM, the TMS and the telematics, and does only what is permitted.
3 systems
answer one dispatcher’s question
NestJS · PostgreSQL · OpenAI API
Construction
Prokta
Tender documents read for a general contractor’s commercial department: an archive of files goes in, a report of terms comes out, every line of it with a page reference.
91%
of the control sample’s critical terms the system found on its own
Python · Qdrant · OpenAI API
Restaurants
Ostera
A chain of 64 restaurants: figures from five systems in one layer, where a question about yesterday’s profit is asked in plain words and answered with the data behind it.
64 restaurants
counting on one set of definitions
NestJS · ClickHouse · OpenAI API
What people usually ask
Contact
Send a description of the task
A reply with the scope, the timeline and a budget estimate comes within 24 hours.
The first call is 30 minutes, with no commitment on your side.
01
You describe the task
Five questions in the form, or a plain email — whichever suits you.
02
We answer within a day
With the scope, the timeline and a budget estimate, based on what you told us.
03
We talk for 30 minutes
To clear up whatever is unclear. It commits you to nothing.