Owner of Storetown Media · since 2012

Jan-Hendrik KummertAI Engineering for E-commerce and small and medium-sized enterprises

I develop LLM assistants, agent-based workflows, and data pipelines that are reliable in everyday use: based on knowledge derived from real data, clear guidelines, and measurable costs. This is backed by 14 years of practical experience with online shops, ERP integrations, and interfaces.

Request a project view projects
  • 3LLM Assistant developed
  • -90%Input costs thanks to prompt caching
  • 552.669Records per feed run
  • 14Years of e-commerce experience

Tornesch near Hamburg · Remote projects throughout Germany

Focus areas

What I'm building – and what's important in the process

AI is worthwhile where it works reliably. That's why data quality, security, and operation are integral to my projects from the very beginning.

LLM Assistant

Consulting, support, and sales via the Claude and OpenAI APIs. Answers are derived from live data and tools, not from the model's memory.

Agent workflows

Multi-stage processes involving web search, structured output, and independent fact-checking. A human makes the final decision where it matters most.

Data & Integration

Feed pipelines, ERP and accounting interfaces, migrations: with validation, quarantine and a way back for every step.

Security & Operations

Hardening, monitoring, and incident response for Magento and WordPress. Deployments with verification steps and automatic rollback.

Projects 2025–2026

Selected works

The examples are either operational or about to launch. All key performance indicators (KPIs) have been measured.

01

Captain Ole, Johnny and Sophie

LLM assistants for consulting, landlord support and sales

storetown-media.de · fewo-nordostsee.de · hairdoshop.de (Pilot) · since 07/2026

Three assistants on the Anthropic API, connected via their own server endpoints – the API key never leaves the server. The knowledge comes from the live database: prices, products, appointments, locations. Sophie works directly with the shop data using five tools and adds items to the shopping cart with just one click from the customer. Ole and Johnny don't save any chats. The model is interchangeable for each assistant via configuration – I test new model generations against the same test questions before switching.

Captain Ole is active on this page – feel free to ask him a question.

  • -90%Input costs for follow-up questions through prompt caching
  • 8Documented testing rounds: finding, rule, retest
  • 9 → 0Mismatches in the test with 32 customer questions
  • Claude Haiku 4.5
  • Claude Sonnet 5
  • Claude Opus 5
  • Tool Use
  • Prompt caching
  • Wordpress
  • PHP

02

Editorial automation with fact-checking

Agent pipeline for event texts

fewo-nordostsee.de · 06–09/2026

For each appointment, two model calls utilize server-side web search: the first researches and writes the output, the second verifies each statement against the source. Output is enforced via a JSON schema; web content is treated as data, not instructions, and "better empty than guessing" is the strict rule. Publication occurs automatically or only after approval.

  • 0,16 $Cost of the first run with a 38-second runtime
  • 2 stagesResearch and adversarial cross-checking
  • 49Municipal portals as verified appointment sources
  • Claude Opus 5
  • Web Search
  • Structured Output
  • PHP
  • Cron

03

Prizavo

Price comparison with feed pipeline and release engineering

Own product · since 09/2026 · Launch in preparation

Offers from affiliate and merchant feeds are streamed, validated, assigned across merchants, and evaluated against the 90-day average. Suspicious deliveries—such as massive price spikes—are quarantined instead of being included in the catalog. Every data version is traceable via checksums and rule versions; a new release is only activated when the build confirms it.

  • 552.669Data records per run from 47 dealers
  • 1.500+Automated tests in CI
  • 112 sfor the complete import on the server
  • next.js
  • TypeScript
  • Node.js
  • SQLite
  • GitHubActions
  • Claude Fable 5.1
  • Claude Opus 5.5
  • GPT-6 Astra (Codex)

04

Incident Response & Hardening

Security for Magento 2 and WordPress

Customer projects and in-house systems · since 09/2025

Forensic analysis of logs and malicious code, cleanup, hardening, and GDPR compliance – in the event of real attacks on online shops. New security rules are tested against 30 days of real-world access before going live. Integrity monitors compare core checksums and automatically restore manipulated files.

  • 19 / 19Attack pattern detected, no legitimate URL blocked
  • 1 dayup to the CVSS-10 patch on four shops
  • 44 → 15Server load after targeted bot defense
  • PHP-FPM
  • Magento 2
  • Wordpress
  • Log forensics
  • Composer Audit

05

AI Visibility (GEO)

So that ChatGPT, Claude & Co. can accurately describe a company

7 web projects · since 06/2026

A curated llms.txt file, a crawler strategy that differentiates between response and training crawlers, and clean, structured data. A proprietary testing tool uses LLM to pose realistic user questions and measures whether a brand appears in the answers – with a fixed daily budget.

AI in e-commerce at Storetown Media
  • 13AI crawler detected in testing tool
  • 88Findings from an audit with eight agents
  • 3 wavesImplementation live on storetown-media.de
  • llms.txt
  • schema.org
  • Open AI API
  • Search Console API
  • MCP

06

Software with test chain

Plugins for WooCommerce and Magento

wordpress.org · Adobe Commerce Marketplace · Open Source

STM Barrierelotse checks checkout processes according to WCAG 2.1 AA and generates the declaration according to § 14 BFSG; a word choice gate in the corporate identity prevents inadmissible claims of conformity. STM Smart Checkout Pro checks the VAT ID number via the EU VIES interface – if it fails, there is no tax exemption.

  • 2Plugins in the wordpress.org directory
  • 4Extensions in the Adobe Commerce Marketplace
  • 195Logic tests in the CI of the Barrier Pilot
  • PHP
  • GitHubActions
  • PHPCS
  • PHPStan
  • axis-core

Operation

Agentic Engineering with clear responsibility

I openly use AI agents: they write code, check each other's work, and conduct extensive research. I am responsible for the architecture, quality, and operation.

  1. 01

    Requirements & Architecture

    Define the goal, data flow, and boundaries. Decisions are logged before the first line of code is written.

  2. 02

    Implementation by agents

    Claude Code is implemented using Claude Fable 5.1 and Opus 5.5 in parallel development phases. Every change is documented as a pull request.

  3. 03

    Independent Audit

    A second model – GPT-6 Astra in OpenAI Codex – tests according to fixed rules. Each finding is first marked with a failed test, then corrected.

  4. 04

    Release & Operation

    Only a human can grant access. After that, monitoring, alarms, and a tested return path take over.

Models in use · As of September 2026

  • DevelopmentClaude Fable 5.1 · Claude Opus 5.5via Claude Code, in parallel in Git worktrees
  • ExaminationGPT-6 Astravia OpenAI Codex as an independent reviewer
  • ApplicationsClaude Opus 5 · Sonnet 5 · Haiku 4.5via the Anthropic API, and also the OpenAI API

I implement new models as soon as they are available and pass my tests. The applications are built to be model-independent – ​​switching is a configuration change, not a complete overhaul.

  • 630+Agent meetings since March 2026
  • 1.216documented decisions, findings and rules
  • 87Pull requests in three weeks – with CI and second review

Guardrails

Six rules that apply to every project

  1. Grounding instead of installments

    Assistants provide answers based on live data and tools. The system simply states what is not verifiable.

  2. Data is not instructions

    Websites, tool results, and inputs are marked as data – the most effective protection against prompt injection.

  3. Measure costs instead of estimating them

    Usage is recorded per call; budgets and rate limits are part of the architecture.

  4. No deployment without a way back

    Verification step before activation, backup before that, automatic rollback if the health check fails.

  5. Findings, period, retest

    Every error becomes a documented rule and is secured with a test.

  6. Data protection from the start

    No chat logs, visible AI labeling, and understandable privacy notices.

Tools

Technologies in use

AI & Agents

  • Claude Fable 5.1
  • Claude Opus 5.5 · Opus 5
  • Sonnet 5 · Haiku 4.5
  • GPT-6 Astra
  • Claude API
  • Open AI API
  • Claude Code
  • OpenAI Codex
  • MCP
  • Tool Use
  • Prompt caching
  • Structured Output

Development

  • PHP
  • TypeScript
  • JavaScript
  • Python
  • SQL
  • next.js
  • React

platforms

  • Magento 2 / Adobe Commerce
  • WooCommerce
  • Wordpress
  • Shopware 6
  • Shopify

Data & Operation

  • MariaDB
  • SQLite
  • Redis
  • Opensearch
  • Varnish
  • GitHubActions
  • Linux

Career

Stations

  1. since 2026

    Focus on AI Engineering

    LLM assistants, agentic workflows and data pipelines in production.

  2. since 12/2015

    Owner & Managing Director, Storetown Media

    Online shops, interfaces and operations for B2B and B2C – in Tornesch near Hamburg.

  3. 2012 - 2015

    Founder, Webdesign-Kummert

    Websites, online shops and digital marketing in Pinneberg.

  4. Academic background

    Master of Arts (MA)

    Christian-Albrechts-University of Kiel; previously social insurance specialist at AOK Hamburg.

  5. Languages:

    German, English, French

    Fluent in English, good in French.

FAQs

Frequently asked questions

Who is Jan-Hendrik Kummert?

Jan-Hendrik Kummert is the owner and managing director of Storetown Media in Tornesch near Hamburg and has been self-employed in e-commerce projects since 2012. His focus today is AI engineering: LLM assistants, agent-based workflows and data pipelines for online shops and SMEs – based on 14 years of practical experience with Magento, WooCommerce, Shopware, Shopify and ERP integrations.

What is Agentic Engineering?

In agentic engineering, AI agents handle the implementation, while a human is responsible for the result. In my project, Claude implements code in parallel workstations, a second model checks according to predefined rules, and automated tests ensure every step is safe – only a human can release the final version. This approach resulted in 87 pull requests for the price comparison site Prizavo in three weeks.

Are you writing the code yourself or with AI?

AI coding agents handle the implementation; I'm responsible for architecture, requirements, review, testing, and release. Every change goes through automated checks and an independent second review as a pull request – only a human can approve it.

Which models do you work with?

In development, I use Claude Fable 5.1 and Claude Opus 5.5 via Claude Code, and as an independent reviewer, I use GPT-6 Astra via OpenAI Codex. The applications run Claude Opus 5, Sonnet 5, and Haiku 4.5 via the Anthropic API and the OpenAI API – small, fast models for chats, and larger ones for research and review. The applications are built model-agnostic: I implement new models as soon as they are available and pass my tests.

How will a AI assistant Connected to the shop and data?

A dedicated server endpoint in the shop or WordPress interface forwards requests to the language model – the API key remains on the server. Knowledge is drawn from the live database, such as products, prices, or dates, or via tools that the model specifically calls: product search, availability, shipping costs. Actions like "add to cart" only occur after a customer click and with server-side validation.

How do you prevent an assistant from inventing things?

The knowledge comes from live data and tools, not from model memory. Limits of statements and explicitly excluded facts are displayed in the prompt, and every weakness identified in a test round becomes a rule with a retest. If an answer is unsubstantiated, the assistant indicates this.

Do you use RAG or a vector database?

Retrieval, yes; a vector database, deliberately not so far. My assistants draw their knowledge directly from the live database, from curated knowledge blocks with keyword selection and via tool calls – this is up-to-date, transparent, and requires no additional infrastructure. A vector search is worthwhile where large amounts of unstructured text need to be searched, such as manuals or support archives.

How do you measure the quality of a AI assistants?

Using standardized test questions instead of gut feeling. Each test round follows the pattern of finding, rule, retest; this is supplemented by targeted tests against prompt injection, topic deviation, and hallucination bait. Knowledge selection is verified by an offline test without API costs: For Captain Ole, this reduced the number of incorrect assignments in 32 typical customer questions from 9 to 0.

Are chat histories saved?

Captain Ole and Johnny do not save chats. Sophie's chat history is stored on the server for two hours to ensure the conversation remains coherent. The use of AI is indicated in the chat window and described in the privacy policy.

How do you handle data protection at AI assistants around?

Data-efficient and transparent: The API key is stored only on the server, inputs are limited in length and number, and the privacy policies of the live systems describe the providers and data transfer. Users wishing to speak to a human are directed to a contact form, live chat, or telephone. Where voice input is available, a separate notification appears before the first recording.

What does it cost to operate a AI assistants?

This depends on the model, the scope of knowledge, and the usage. In one project, an answer using the small model cost less than half a cent; prompt caching reduced the input costs for follow-up questions by approximately 90%. Budgets and rate limits put an upper limit on costs.

Do you also work remotely and on a project basis?

Yes. I work from Tornesch near Hamburg; projects are run remotely throughout Germany, on-site meetings are possible by arrangement.

You are planning to conduct an AI application or a data pipeline.

Briefly describe your request. I will get back to you within one business day – the initial consultation is free.

Request a project jhk@storetown-media.de

Or call us: 04122 4084792 · More about Storetown Media