[ 00 ] — AI Engineering · Data Architecture

Working AI systems.
Not slides.

We design and ship agentic systems, LLM platforms and production data pipelines. Reproducible, auditable — with the IP on your side.

9+
years in Data Science / ML
6,231
competencies in our taxonomy (3 models, voting)
414
AI glossary entries (22 models, 3 rounds)
models in the multi-model consensus method

[ 01 ] — SERVICES

Four ways we deliver — plus an ongoing engagement.

A named scope, a fixed timeframe, a concrete outcome. Pricing after a conversation about your data.

PAKIET 01 · 4–6 weeks

RAG / Agent Proof-of-Value

For companies that want proof of value before investing in a platform.

You get
  • A working demonstrator on your data
  • A response-quality evaluation dashboard
  • A go/no-go decision report

We define the business metric in week 1 — so your pilot doesn’t become another statistic among the 95% of projects with no ROI. If the results don’t defend the metric, you get an honest report on why and what’s next.

Get in touch →
PAKIET 02 · 2–3 weeks

LLM Observability Sprint

For teams running LLMs in production with no visibility into quality and cost.

You get
  • Observability (Langfuse / OpenTelemetry)
  • LLM-as-a-Judge evaluation with an audit trail
  • Cost per request

An audit trail and evaluation records aligned with EU AI Act requirements (high-risk system obligations from 08.2026). We deliver the engineering side of compliance — we leave interpreting the regulations to lawyers.

Get in touch →
PAKIET 03 · 2 weeks
Most common starting point

AI Architecture Blueprint

For boards and CTOs ahead of an investment decision.

You get
  • A target-architecture document
  • An implementation backlog
  • A TCO estimate

An independent audit: you get the document, the backlog and the estimate — you can build with anyone. We don’t upsell ourselves an implementation “on the side”.

Get in touch →
PAKIET 04 · from 3 weeks

PoC-to-Production

For DS teams with notebooks that need to deliver repeatably.

You get
  • Reproducible pipelines (Kedro / MLflow)
  • Data validation and CI/CD for ML
  • Full auditability

From pilot to production: nearly half of PoCs never get there — we close that gap.

Get in touch →
PAKIET 05 · retainer, 2–4 days / month

AI Engineer on your team

For companies that need a senior GenAI engineer on an ongoing, flexible basis.

You get
  • Ongoing access to a senior AI engineer
  • Architecture and code reviews
  • Team mentoring

Max. 2–3 clients in parallel — exclusivity instead of agency scale. Piotr builds himself, not just advises.

Get in touch →

Not sure where to start? Book 30 minutes →

[ 02 ] — FOR WHOM

Who we work with

01

Mid-market and scale-ups with a stalled pilot

You have a GenAI PoC that hallucinates or doesn’t deliver ROI. We take it to production or kill it honestly.

02

Software houses and integrators

You added AI to your offering, but you lack a senior GenAI engineer at production level. We bring the AI layer under your brand — project subcontracting, not staff leasing.

03

Companies with high-risk systems (EU AI Act)

HR-tech, scoring, fintech — you need the engineering side of compliance: logging, documentation, evaluation and an audit trail.

[ 03 ] — PROCESS

Incremental. Payment on milestone acceptance.

01

Talk

A free consultation. We define the problem and the success criteria.

02

Blueprint

Architecture, milestones, estimate. You know what you’re paying for.

03

Build

Every stage ends in a working product. You accept it — you pay.

04

Handover

The IP, documentation and reproducible pipelines stay with you.

[ 04 ] — WORK

Selected work

Independent Datarmination work — our own methods and tools.

Professional competency taxonomy

Problem
No coherent, searchable competency structure from job postings.
Approach
A 6,231-competency taxonomy across 21 categories, built via multi-model consensus — 3 models, majority voting.
Stack
SFIA v9 / ESCO / O*NET / SCOR / BABOK, Python
Value
Auditable, hierarchical taxonomy with transparent voting, mapped relations and semantic search.

AI glossary 2022–2026

Problem
Fragmented, inconsistent AI terminology with no links between concepts.
Approach
414 entries synthesized from 22 models across 3 rounds, with a 104-relation concept graph and source verification.
Stack
multi-model synthesis, knowledge graph (Graphviz), Python, React
Value
A reproducible reference glossary: 358 verified sources, term-maturity ratings, evolution chains and interactive visualizations.

LLM tooling

Problem
No visibility into LLM project progress and costs.
Approach
LLM project tracker (FastAPI + React, interactive milestone graph) and an LLM-conversation analysis pipeline → SQLite → faceted taxonomy tagging and Obsidian export.
Stack
FastAPI, React (React Flow), SQLite, Python
Value
Central visibility into and organization of work with LLM models.

Deep-research orchestration

Problem
A single research model hallucinates and misses sources.
Approach
Parallel runs of multiple deep-research models, result triangulation (a fact confirmed by ≥2 tracks = certain), anti-hallucination regime (URL citations, explicit uncertainty), merge into a consensus report.
Stack
orchestration meta-prompt, multiple DR models, triangulation
Value
A reusable research template with documented sources and divergences.

[ 05 ] — FIRST CALL

How the first call works

  1. 0130 minutes, no commitments.
  2. 02We define the problem and the success criterion.
  3. 03You leave with a scope draft and a package recommendation.

[ 06 ] — FAQ

Frequently asked questions

Why no pricing?+

Because a fair quote depends on your data and scale. We give a fixed price after a 30-minute scoping — upfront, not “as we go”.

What if the project outgrows one person?+

The scope of each stage is clear and finite. I build the core myself; for overload I have a consortium model and trusted partners.

Whose IP is it?+

Yours. The code, prompts and pipelines in full — we deliver them reproducibly, and you run them yourself.

How do we settle up?+

On acceptance of each milestone. You don’t pay for a stage you haven’t accepted.

NDA and data security?+

NDA as standard, before any conversation about data. We work on your infrastructure wherever data can’t leave it.

What happens after the Proof-of-Value?+

Either hardening to production (PoC-to-Production), or an ongoing engagement on a retainer model.

How do you measure LLM response quality?+

A test set + LLM-as-a-Judge + metrics defined upfront. Without that, “it works” is an opinion, not a fact.

[ 07 ] — CONTACT

Contact

Let’s talk about your project — from feasibility analysis to a working demonstrator.

I reply within 24h on business days.

About Datarmination

Datarmination is run by Piotr Pytlarz — an AI engineer and data architect with 9+ years in Data Science / ML. He held a key role in the internal ML/AI team of one of Poland’s largest IT consulting firms (7,500+ specialists), designing agentic RAG systems and LLM platforms for enterprise use. Previously Data Science Team Leader and Data Scientist Lead. Datarmination delivers projects independently and in consortia.

piotr@pytlarz.pl

+48 572 759 900

LinkedIn ↗


Datarmination Piotr Pytlarz

NIP 9680843592

ul. Chłapowskiego 13, 02-787 Warszawa