Applied AI engineering studio

We build the AI systems behind your product.

Ozon Data Services designs, ships and runs production LLM software: retrieval, agents, evaluation and the infrastructure that keeps it fast, safe and measurable. Senior engineers, shipping in weeks, not slideware.

Start a project See our work

40+AI systems shipped

<120msmedian inference

7countries served

ozon-llm-ops · production LIVE

Requests / min

3,418

▲ 12.4%

p95 latency

118ms

▼ 22ms

Cost / 1k

€0.31

▼ 18%

Throughput & latency · last 24hreq/slatency

PASSeval: retrieval-quality0.94

PASSeval: answer-faithfulness0.91

PASSguardrail: pii-redaction1.00

Eval suite green

142/142 checks

Auto-scaled

0 cold starts

What we do

From prototype to production AI, owned end to end.

We are the engineering team companies bring in when an AI idea has to become real software that people depend on every day.

Custom AI development

Bespoke LLM applications built around your data, workflows and constraints, not a generic template.

Model selection & fine-tuning
Retrieval-augmented generation
Agentic workflows
Human-in-the-loop review

LLM integration

We wire models into the products you already run, with clean APIs and predictable behaviour.

API & SDK implementation
Prompt & context engineering
Streaming & tool-calling
Fallbacks & rate control

AI infrastructure

The unglamorous layer that makes AI reliable: deploys, scaling, observability and cost control.

Cloud & edge deployment
Autoscaling & caching
Tracing & monitoring
Cost optimisation

Analytics & insight

Turn messy operational data into models and dashboards your team actually reads.

Data pipelines
Predictive models
Business intelligence
Custom dashboards

Security & compliance

Guardrails, redaction and audit trails so AI can touch real data without becoming a liability.

PII detection & redaction
Access control
Evaluation & red-teaming
GDPR-ready processing

Not sure where AI fits yet?

Book a scoping session. We map the highest-leverage use case and give you a plan with real numbers, no obligation.

Talk to an engineer

How we work

A small senior team, embedded with yours.

No offshore black box. You get named engineers who understand your domain, ship in short cycles, and hand over systems you can actually maintain.

Weeks, not quarters
A working prototype in the first sprint, in front of real users fast.
Measured, not vibes
Every release ships with an evaluation suite so quality is a number, not an opinion.
Yours to keep
Clean code, documentation and a handover so your team owns what we build.

Selected work

Systems in production, results that moved.

−63%support handling time

Ozon built a retrieval assistant on top of our five years of support tickets and product docs. It drafts every reply our agents send now, grounded in our own content, and the evaluation harness they set up means we can change prompts without holding our breath. Handling time dropped by nearly two-thirds in the first quarter and our CSAT actually went up.

Martina Kovacs

Head of Support, Northwind SaaS

4.2×faster document processing

We drowned in scanned contracts. The Ozon team shipped an extraction pipeline with human review that reads, classifies and structures every document into our system, with a confidence score and a full audit trail. What took a team two days now runs in hours, and legal trusts it because they can see exactly why every field was filled the way it was.

Daniel Petrov

COO, Meridian Capital

For developers

Your data, one endpoint, agent-ready.

Everything we build ships with a clean API and a Model Context Protocol server, so your own tools and AI agents can query it securely, with scoped, logged, read-controlled access.

Read the docs Get an API key

quickstart.ts

// Connect an agent to your Ozon data via MCP
import { OzonClient } from "@ozon/mcp";

const ozon = new OzonClient({ apiKey: process.env.OZON_KEY });

const answer = await ozon.query({
  tool: "knowledge.search",
  input: "Q3 churn drivers by segment",
  grounded: true
});

console.log(answer.citations); // fully sourced

About Ozon

A Bulgaria-based studio, building for teams worldwide.

We are a compact team of engineers and applied researchers in Plovdiv who care about one thing: AI that survives contact with real users. We have shipped assistants, extraction pipelines, agents and the infrastructure under them across support, finance, legal and operations.

We work in English, keep European working hours, and are comfortable inside regulated environments where every decision has to be explainable.

Work with us