Roland Polczer

AI Engineer · Systems integration & custom builds

Modern AI, fluent in your systems.

AI projects in mid-sized companies tend to fail at the same seam: the people who understand the models don't understand the systems the business actually runs on, and the people who run those systems don't understand the models. I've spent 25 years on the enterprise side — ERP, billing systems, databases, interfaces, data pipelines — and the last years building agentic AI that connects to real tools and real data. I work at that seam, for the German Mittelstand.

See the work ↗ Get in touch ↗
25+ yrs mission-critical data100% on-prem inference~$0/mo autonomous platform
Scroll

What I do

01

Custom builds & system integration

The core of the work: agentic systems shaped around your workflows and wired into what you already run — ERP customization and interfaces, databases, file shares, internal APIs. Tool calling, orchestration patterns and automation pipelines, with evaluation and guardrails so you can tell whether it actually works. Delivered in short increments, against real performance constraints.

Tool callingOrchestration patternsERP interfacesEvaluation & guardrailsIncremental delivery
Services ↗
02

Document & data processing

Your information arrives as Excel workbooks, PDFs, scans and email follow-ups. I build conservative pipelines that turn all of it into structured, validated data — schema gates, confidence thresholds, human review where it matters, and a full audit trail. Output lands in exactly the format your downstream system needs.

ExcelPDFScansEmailSchema-validated
Services ↗
03

Private & EU-compliant AI

Two ways to keep company data where it belongs: fully on-prem inference on your own hardware, or Azure AI Foundry EU Data Zone deployments, where prompts and responses are processed inside Microsoft's EU Data Boundary. No content used for training, no per-token surprises, and a clear answer for your data protection officer before they ask.

On-prem inferenceEU Data ZoneGDPRNo training on your data
Services ↗
04

Training & board-level advice

Hands-on training for the teams who will use these systems, and clear briefings for the people who have to decide about them. For boards and management: what to expect from AI over the next planning cycle — adoption scenarios, organisational impact, risk and governance, and concrete next steps. Including when the honest answer is 'not yet'.

Team trainingBoard & leadership briefingsAdoption scenariosRisk & governanceNext steps
Services ↗

Selected work

  • 01

    CrewTaxi

    Document Intelligence

    Crew-change information for maritime transport arrives as multi-sheet Excel files, PDF crew lists, scans and last-minute emails. CrewTaxi ingests all of it, validates every extracted fact against JSON Schema, merges updates into one auditable record per person, and exports a clean CSV for the booking system. Deterministic extraction where layouts are known, LLM extraction where they aren't — nothing enters the system without passing a validation gate.

    Schema-gated LLM outputField-level provenanceAppend-only audit logIdempotent batch runs
  • 02

    In-house AI for mid-sized companies

    Corporate Enablement

    A set of internal tools built for mid-sized companies: a document translator that preserves layout across PPTX, DOCX and PDF; retrieval-augmented assistants grounded in the company's own documentation; deep-research agents that compile sourced briefings; and a real-time speech-to-text translator for multilingual meetings. Everything runs on Azure AI Foundry EU Data Zone deployments, so prompts and responses are processed inside Microsoft's EU Data Boundary and company content is never used to train the models.

    PPTX · DOCX · PDFRAG assistantsDeep-research agentsLive meeting translationEU Data Zone deployments
  • 03

    OpenClaw Seismograph

    Autonomous Systems

    An agentic news-intelligence platform that monitors the AI-agent ecosystem daily, scores model predictions monthly, and publishes bilingual intelligence reports — fully automated on GitHub Actions and Cloudflare Workers. No database, no servers to maintain, near-zero running cost. Built for solo-operator maintenance: it runs itself.

    GitHub ActionsCloudflare Workers~$0/month run costZero-maintenance design
    openclaw.lu ↗
  • 04

    Private content factory

    On-Prem AI

    A private content factory that produces and operates fully automated social media channels end to end — research, script, image, video, voice and music, then publishing, metrics and the planning of what comes next, all from one orchestration layer. Postgres-backed job queue, QA gates on frames and video, analytics feeding back into the next plan. Every model runs on-prem, so there are no per-token API bills and no third-party inference anywhere in the pipeline.

    Multi-agent orchestrationQA gates100% on-prem modelsPublishing daily
    See it publish: @SprachKino ↗

Showreel

Fully automated content pipelines.

Two of the channels the factory runs. Research, script, image, video, voice and music all run on-prem — no per-token API bills, so each additional video costs next to nothing, and nothing is uploaded by hand.

Grid of SprachKino video thumbnails: cinematic scenes teaching German phrases

SprachKino

German through cinematic 30-second micro-scenes.

72 videos · fully autonomous
Grid of Too Close to Nature video thumbnails: a host standing unwisely close to wild animals

Too Close to Nature

Real wildlife facts, explained from a terrible distance.

57 videos · fully autonomous

Stack

The stack behind the work.

Everything here is in daily production use.

Agents & orchestration

Model Context Protocol (MCP)Custom Python agentsMulti-agent workflowsEvaluation & guardrailsPersonas · memory · routingHermes AgentDeerFlow

Inference & LLMOps

vLLMllama.cppLiteLLM routingUnsloth fine-tuningLoRA trainingAzure AI FoundryOpenRouterGroqSambaNovaGoogle AI Studio

Data & retrieval

PostgreSQLpgvectorSupabaseOracle · PL/SQLMS SQLFirestoreSearXNGCrawl4AIETL/ELT · Azure Data FactoryStar schema · Data Vault 2.0

Media & infrastructure

ComfyUIDiffusion image & video modelsVoice design · TTSDockerGitHub ActionsCloudflare WorkersTailscale meshUbuntu · WSL

Both worlds

Portrait of Roland Polczer

Enterprise reality: telco billing and rating at Deutsche Telekom and T-Mobile, IFS ERP rollouts, migrations and reporting, cloud data warehouses, high-volume Oracle pipelines. I know how an order becomes an invoice, where master data quietly goes wrong, and what a cut-over weekend costs. Modern AI: agents, orchestration, retrieval, evaluation and guardrails — built and operated daily, not read about. Most people in this field have one of those halves. The interesting work happens where they meet, and that is where I sit.

  • Based in Lienen, Germany
  • Working languages: English, German, Hungarian
  • IFS ERP · Oracle · MS SQL · Data Vault
  • Led teams of up to 12
  • B.Sc. Computer Engineering · B.Sc. Finance & Accounting
  • Top IT Worker of the Year — T-Mobile, 2002

Writing

I write about where AI development is actually heading — hands-on, not hype.

All posts on Substack →

Contact

Available for new projects

Get in touch.

The fastest way to reach me is email — I read everything myself. If you have documents that should be data, or an AI idea that needs to survive contact with production, describe it in three sentences and I'll tell you honestly whether I can help.

Reply within 24 hours · Free intro call · Lienen, NRW — on-site in Münsterland & remote