Services

AI solutions

Assistants, search over your own data, routine work handed to an LLM.

Language models start earning their keep where there is a repeating operation on text: support answering the same questions, a manager triaging inbound requests by hand, company knowledge spread across hundreds of documents nobody can search. That shape of problem is solved by retrieval over your own data plus a model that is only permitted to answer from what it retrieved. A project whose value is stated as "we need AI", with no named operation it replaces, is an expensive demo — and I will say so on the first call.

What the work covers

An assistant over your knowledge base: document indexing, retrieval, answers that cite their source (RAG)
A model integrated into an existing product: request classification, field extraction from mail and documents, draft replies
The whole chain automated: request → parse → task in the CRM → notification to whoever owns it
A quality harness: test questions with reference answers, so swapping a model or a prompt can be measured instead of sensed
A data perimeter: what leaves for the model provider, what never leaves, where request logs are kept and for how long
The honest non-model alternative, when rules and full-text search solve the same task for less

Who this fits

A fit: there is a volume of same-shaped text work and a named person doing it by hand today
A fit: you own the data — documents, a database, correspondence — and it is solid enough to answer from
A fit: you accept that the model gets things wrong, and the process has a place for a human to check it
Not a fit: you need guaranteed accuracy. A model produces a probabilistic result; where an error costs money, health or legal standing, you need a different tool or mandatory review
Not a fit: the data cannot leave your perimeter and there is no budget for a local model and the hardware under it
Not a fit: the goal is being able to tell an investor or a conference that the product has AI in it

Stack

Python
FastAPI
PostgreSQL
pgvector
TypeScript
OpenAI-compatible API

What moves the price

The state of the data: a clean database, or an archive of PDFs and scans that first has to be recognised and normalised
Corpus size and how often it changes: a quarterly re-index and a daily stream of edits are not the same price
Privacy requirements: an external API, a Russian provider, or a model running on your own hardware
How many integrations surround the model: one form on a site, or a chain through CRM, mail and a messenger
Whether a formal quality harness is required and maintained after launch — that is separate, ongoing work
Who pays for tokens: running the model is not part of the build price. The provider account is opened in your name and billed to you directly, so the spend is visible to you in real time and never passes through the studio with a markup

What you keep

Sources, prompts and the indexing configuration in your repository — the prompt does not live in a contractor's head
The test question set and the last run's results, so any later edit has something to be compared against
A written data perimeter: what goes to the provider, what stays with you, where logs sit and when they are deleted
Instructions for switching model or provider, so the solution is not locked to one vendor
Keys and billing on your own accounts

Packages

Oracle

An AI integration: an assistant over your data, or a processing pipeline.

from $14,000

Timeline: 4–8 weeks
Payment by stage: 40 / 40 / 20

Included

An assistant over your knowledge base that cites a source in every answer
A document and data processing pipeline
Explicit limits on what the model can reach, and an access log
Answer quality measured against your own examples, not against a demo
The model provider can be swapped without rewriting the application

Not included

Training or fine-tuning models of our own
Model-provider tokens and subscriptions — on your own accounts and at your own cost
A guarantee that a generative model is factually correct
Sending personal data to foreign providers without a lawful basis

Discuss this package

Retainer

Ongoing

Ongoing support: a block of hours every month and an agreed response window.

from $2,000/mo

Timeline: minimum 3 months

Included

Twenty hours a month for fixes and development
Incident response within two working hours — weekdays, 10:00 to 19:00 Moscow time
Dependency updates and triage of new vulnerabilities
Backups watched, and restores tested
A monthly report: what was done and where the hours went

Not included

Rolling unused hours into the next month — the fee reserves capacity rather than buying a pack of hours; for one-off work there is Hour
Round-the-clock on-call
A new product from scratch — that is a project of its own
Hosting, licences and third-party services

Discuss this package

Hour

Ongoing

Hourly work: one task, a consultation, a review of someone else's code.

$120/hr

Timeline: minimum block 4 hours

Included

A four-hour minimum block
A time report describing what the hours went on
A review of existing code and infrastructure, with written findings
A consultation on architecture, on the stack, or on someone else's quote

Not included

A whole project — past a few days of work a package costs less
Scheduled on-call — that is Retainer
Availability on a given day without booking it in advance
Work that starts before the task is written down

Discuss this package

Discuss a project All services