Corexy
S/06 — ai and machine learning

AI & LLM Integration for Business Systems

AI integration means wiring a language model into your own data and systems: retrieval over internal documents, an assistant inside support, classifiers and extraction pipelines that replace manual reading. We treat it as ordinary engineering — one process, one retrieval layer, one number that has to move — and your data stays where you decide it stays.

Discuss your taskScope in 1 day, reply within 2 hours

What this solves

Your data cannot leave the perimeter

The blocker in most AI deals is not capability, it is where the text goes. We design around it from the first call: API terms that exclude your traffic from training, or an open model on your own GPU with nothing leaving the network.

Answers buried in internal documents

Policies, contracts, spec sheets and a decade of wiki pages. We build retrieval that finds the right passage and an assistant that answers with a citation to the source file, so a wrong answer is traceable instead of mysterious.

Documents typed into systems by hand

Invoices, purchase orders, claims, onboarding forms. An extraction pipeline pulls the fields, validates them against your own reference data, and routes anything failing a check to a human queue instead of guessing.

Support repeating itself at volume

Tier-one tickets that are the same twelve questions. We start in draft mode — the model proposes, the agent sends — measure the edit rate for a few weeks, and only then open up the parts that earned autonomy.

A proof of concept that never shipped

Notebook demos are easy. The hard part is retries, evaluation, cost control and the week your provider deprecates a model, so we ship this as a service with logs, tests and a rollback path like anything else we run.

How the work runs

  1. Scope one process, agree the metric

    1–2 days

    We look at a single workflow: volume per day, minutes per item, where the time actually goes. One number is agreed as the thing that has to move, because everything after this is judged against it.

  2. Prototype on your real data

    5–7 days

    We take a real sample — 100 to 300 documents or conversations — and build a working prototype against it. This is where you find out whether a model solves your case at all, before the main budget.

  3. Build the production pipeline

    10–15 days

    Ingestion, chunking, retrieval, an evaluation set, guardrails, and integration with your CRM, helpdesk or app. Deployed in your cloud or ours, every request logged, every answer traceable to a source.

  4. Run it beside a human

    5–7 days

    The model drafts, your team reviews and corrects. We track edit rate and handling time against the metric from step one, and tune retrieval on the failures that actually happen rather than imagined ones.

  5. Handover and cost model

    2–3 days

    Source code, documentation, and a plain spreadsheet of monthly token and hosting cost at your volume. One month of warranty: bugs in delivered functionality are fixed free.

Stack

Model choice follows the data constraint, not the leaderboard: cloud models where text may leave, open weights on your hardware where it may not. Most of the effort goes into retrieval and data plumbing, which is why we treat this as a database problem first.

  • OpenAI API
  • Anthropic Claude API
  • Google Gemini
  • Qwen / Llama (self-hosted)
  • Ollama / vLLM
  • PostgreSQL 16 + pgvector
  • Qdrant
  • Python + FastAPI
  • Node.js / TypeScript
  • Docker

What it costs

Proof of concept

from $900 / from 90 000 ₽

One process, your data, a working prototype and a measured result — roughly 30 hours. You get accuracy numbers and a cost per item before committing to a build, and the prototype code is yours either way.

Production build

from $2,400 / from 240 000 ₽

The full pipeline: ingestion, retrieval, evaluation set, guardrails, integration and an admin surface for the knowledge base — roughly 80 hours. Price moves with the number of sources, languages and self-hosting requirements.

Ongoing engineering

from $400 / from 40 000 ₽ per month

A dedicated engineer on retainer: knowledge base updates, failure review, retrieval and prompt tuning, token spend monitoring, and migration work when a model is deprecated. Hours beyond the package are billed from $30.

Prices are indicative and do not constitute a public offer. The final cost, scope and timeline are fixed in a contract or invoice after we have discussed the task.

Related case study

The closest project is Hook — a Telegram travel bot with a CRM admin panel and an AI lead channel. It runs 12+ products over 13 database tables, with Gemini on the cloud side and a locally hosted Qwen where a request should not leave the server. That split is the same pattern we propose when part of your data is sensitive and part of it is not.

Read the Hook case

Frequently asked questions

How much does it cost to integrate AI into our product?
A proof of concept on one process starts at $900 / 90 000 ₽, a production build at $2,400 / 240 000 ₽. We bill hourly from $30, quoted after a free scoping call, with the standard 50% upfront and 50% on acceptance. Model and hosting costs sit outside that and are estimated at your volume before you commit, not discovered in month three.
Will our data be used to train someone else's model?
No, and it is worth being precise about why. API terms at OpenAI, Anthropic and Google exclude API traffic from training by default — that is a contractual guarantee, not an architectural one. If your security review needs an architectural guarantee, we run an open model on your own hardware, where no request leaves your network at all.
Do we need RAG or fine-tuning?
Almost always RAG, and we will argue the point rather than hedge. Fine-tuning changes how a model writes, not what it knows: it cannot learn a price that changed this morning, it has to be redone whenever your data does, and it tends to make wrong answers more confident. Retrieval keeps facts in a database you control and updates in seconds. Fine-tuning earns its place for format and tone, or a narrow high-volume classifier.
Which model do you use — OpenAI, Claude, or open-source?
Whichever survives evaluation on your data; we are not a reseller for anyone. In practice cloud models win on reasoning-heavy work, open weights like Qwen win when data cannot leave, and a small classifier beats both on cost for high-volume routing. We keep the provider behind an abstraction so a swap is a config change, because you will make one within a year.
How do you stop the model from hallucinating?
By constraining what it may answer from, not by asking it nicely in a prompt. Responses are grounded in retrieved passages, every answer carries a citation to its source, and "I don't know" plus a handoff is a valid outcome. We also keep an evaluation set of real questions with known answers, so a prompt or model change that degrades accuracy fails before release.
Can the AI run on our own infrastructure?
Yes — self-hosting is a standard option here, not an upsell. Open-weight models run on your GPU or private cloud through Ollama or vLLM, with PostgreSQL 16 and pgvector for retrieval, all containerised. Expect a real quality gap against frontier models on hard reasoning, and close to none on document retrieval and classification, which is most of this work.
What are the ongoing token and hosting costs?
For a typical document assistant, cloud model usage lands in the tens of dollars a month at low volume and scales roughly linearly with traffic; self-hosting trades that for a GPU instance you rent monthly. We put real numbers at your volume into a spreadsheet during the proof of concept, together with the caching and model-tiering options that bring them down.
How do you measure whether the AI solution actually works?
One metric agreed before we start, measured the same way afterwards. Usually it is handling time per item, the share of items closed without a human, or the edit rate on drafted answers. An evaluation set of real cases keeps accuracy a number rather than an impression. If the pilot misses the target, we tell you that instead of proposing phase two.

Bring one process, we'll tell you if a model solves it

Send the workflow and a sample of the data. Within a day you get a straight answer on whether this is an AI problem, what the pipeline would look like, and what the proof of concept costs. Replies within 2 hours in business hours.

Discuss your task
AI & LLM Integration for Business — Corexy