Knowledge retrieval
Your documentation, contracts, and internal wikis become answerable. People ask in plain language and get the answer with the source attached.
AI agents and retrieval systems for B2B service companies
Your team is answering the same questions every day.
We build AI agents and retrieval systems that handle it instead — shipped in two weeks, with a quality threshold written into the contract.
Book a 20-minute call See what we've built
Ask this one anything about the Stripe API docs. It's the same stack we'd build for you.
Example — demo not live yet
What's the difference between a PaymentIntent and a SetupIntent?
A PaymentIntent tracks a payment you are collecting now — it moves through confirmation, any required authentication, and capture of a specific amount.
A SetupIntent runs the same authentication flow but saves a payment method for future payments without charging it, which is what you want before billing a customer later or on a schedule.
Request failed The demo endpoint didn't answer. That's on us, not you.
What we do
Inbox and intake agents
Incoming requests get read, classified, and routed. The routine ones get answered. Your team sees only what needs a human.
Integration into what you already run
The systems above connect to your CRM, ticketing, and databases. No new tool for your team to learn.
Built on LangGraph, Qdrant, and the OpenAI and Anthropic APIs. Your data stays in your infrastructure and is never used for model training.
How it works
-
01Discovery call20 minutes
You describe the process. We tell you whether it's worth automating. Sometimes the answer is no.
-
02Scope and eval set2 days
We agree on what “working” means before we build: a set of 50–100 real questions from your side, and the pass threshold.
-
03Working prototypeweek 1
On your real data, running where you can click it. Not slides.
-
04Delivery and acceptanceweek 2
We run the eval set with you. It passes the agreed threshold or we keep working, at no extra cost.
-
05Supportongoing, optional
Monitoring, updates, and changes as your data changes.
Most AI projects have no definition of done.
The system answers well on some questions and badly on others, and nobody agreed in advance which ones matter. The project drifts, the invoice grows, and everyone gets tired.
We do it differently. Before we write a line of code, you give us 50–100 real questions from your business. We agree on the scoring scale and the pass threshold, and both go into the contract. Acceptance is running that set together — not an opinion about whether it feels good.
If it doesn't pass, we keep working. That's our risk, not yours.
What the manual version already costs you
What it costs
Fixed price, agreed before we start. No hourly billing, no surprises at the end.
Audit
$900
3–5 days
- What you get
- Process map, automation candidates ranked by hours saved, effort and cost estimate for each.
- Who it's for
- You know something is wasting hours, but not what to automate first.
Build
$4,000–5,000
2–3 weeks
- What you get
- One agent or retrieval system, built on your data, with acceptance criteria agreed in writing before we start.
- Who it's for
- You know the process. You want it working in two weeks.
Extended
from $12,000
4–6 weeks
- What you get
- Multiple connected systems, custom interface, integration into your existing stack.
- Who it's for
- The process spans several systems and teams.
The audit comes off the build. If you go ahead with a project within 30 days, the full audit fee is deducted from it. You're not paying twice.
Not included in any package: model inference costs — you keep your own OpenAI or Anthropic account and we'll estimate the monthly spend before we start. Hosting and third-party licences are also yours.
Support after delivery: $500–800/month. Monitoring, data updates, small changes. Optional, cancel any time. First 30 days after delivery are covered free — that's warranty, not support.
What we've built
Retrieval over 1,400 pages of API documentation
Built as a reference implementation. Evaluated on 80 questions written before the build.
- Correct
- 87%
- Partial
- 9%
- Wrong
- 4%
- Median response
- 1.4s
Who you're working with
Maksym Kononenko
I build the systems. 1.5 years shipping LLM applications in production: retrieval pipelines, multi-agent workflows, vector search.
I'm a one-person studio in Ukraine. That means you talk to the person writing the code, and there's no account manager in between.
Questions you're about to ask
How do we know it actually works?
You define the test set before we build. Acceptance is passing it. Details above.
Where does our data go?
Into your infrastructure. Model providers we use don't train on API data by default, and we'll sign a DPA. We can list every subprocessor involved.
You're in Ukraine — is that a risk?
Fair question. We run on backup power and satellite internet, and all work lives in European cloud infrastructure rather than on local machines. Everything is built in your own repository and your own cloud account, documented as we go, so any competent engineer can pick it up without us. The industry here has delivered without interruption for four years.
Who pays for the API usage?
You do, on your own account. We'll estimate the monthly cost before starting so there are no surprises.
What happens when the model gets deprecated?
Covered by the support plan, or quoted separately. We'll tell you which before you sign.
Can you just fix what we already built?
Often yes, and it's usually cheaper than rebuilding. Send us what you have.
You've got a process that eats hours. We'll tell you in 20 minutes whether it's worth automating — and if it isn't, we'll say so.
Or write to max@kaleth.dev.
Don't want a call?
Three fields. We reply within one business day.