Skip to content
Ben Dangelmayr

SAP & AI Consultant · AI engineering

Ben Dangelmayr

I build software and AI agents with coding agents, from architecture to production, including the checks that agent work has to pass.

SAP AI ambassador · AI competence team · Heidelberg

Proof: my boot check caught a server that could not start, while 496 agent-written tests stayed green.Verification

Not another template. The work, and the checks behind it.

Selected work Contact GitHub

Legal notice · Privacy

A drawing board on a long wooden workbench by a window in the evening; a printed template web page, crossed out in red marker, lies crumpled in a felt waste-paper basket below the bench. A cup of tea steams beside the board, a kettle next to it.

Selected work

Not every object is a project. Some are just for fun, like the seals and the cup of tea: click them.

Not another template. The work, and the checks behind it.

Legal notice · Privacy

A developer’s study in warm evening light: a long wooden workbench in front of a pegboard with tools, framed seal photos, a wall clock and shelves; a laptop, a printer and other work stations on the bench, a PC tower and plants below.

Tests written by coding agents

Sabotage probes for AI-generated tests

Own project, built with coding agents Part of the project’s verify gate 2026

496 / 496 green boot check red

Tests written by coding agents often pass whether the code works or not.

  • Probe: I plant one known fault (a server on an address nobody can reach) and run the full check.
  • Unit tests: all 496 stay green.
  • Boot check: starts the real server, calls /health: no answer, exit code 1.
Details
  1. Type check reads every file, runs nothing. Pass.
  2. Unit tests: 496 tests, in-process, no network. 496 of 496 pass, with and without the fault.
  3. Boot check builds and starts the real server and calls /health. With the fault: no answer.
  4. Exit code: 1 with the fault, 0 without.

Without the probe, the unit tests look the same either way: green.

Run
2026-08-07, npm run verify

Log reconstructed from the project’s scripts, project name removed. The clean run is reconstructed from the checks the gate performs. No timings are shown because none were recorded.

A modern laptop on the workbench shows the check log: type check passes, 496 of 496 tests pass, the boot check is red with exit code 1. Next to it a sabotage switch, three status lights and a multimeter.

Anonymized: employer, platform, agent name

LLM agent for CVs, in production

Built the agent, fixed it after launch Sales and consultants, on real bids

Same structure every run, fewer invented metrics

Setup: an agent for sales bids

An agent on a hosted AI platform drafts consultant CVs for sales bids. The platform’s limit: no check step before a user sees the result.

Output: a CV draft per bid

Sales and the consultants use the drafts on real bids. The place for a check stays empty.

Failure: invented metrics, unstable versions

The agent invented metrics that appear in no project data, such as increased efficiency by 50%. Asking twice for the same CV gave very different versions.

Fix: guideline, lower temperature, fixed JSON

  • Guideline from sales in the knowledge base
  • Temperature lowered, output fixed to one JSON structure
  • Since then: same structure every run; invented metrics less often, not never. A check step is still missing.
Details

Configuration only: the platform lets you set the system prompt, model, temperature, integrations, knowledge base and JSON output, and runs no code of my own. The guideline uses one of the best real CVs as its example. That invented metrics came up less often is an observation, not a measurement.

A light desktop printer prints CV sheets with photo fields and bars; a guideline binder sits in its knowledge-base tray, the temperature dial on the printer points to green. Beside it a dry stamp pad with an unused stamp.

SAP BAdI extension of a standard interface

Extending SAP without touching SAP

Maintained the SAP side, built the extension In production, signed off by accounting 2025

Every card transaction keeps its own posting line

  • Problem: a standard interface condensed the card lines of an expense report into one posting line, so one accounting team had to reconcile by hand.
  • Fix: an extension at a supported extension point runs before condensing; the standard program stays unchanged.
  • Effort: a few lines of code. Most of the work was proving the extension point runs early enough with the data it needs.
  • Sign-off: accounting confirmed it in the test system in writing, then it went to production.
Details

The extension runs only for that team’s companies. A setting was not enough: condensing is switched per line type, for all companies at once.

A flat, dark booking terminal with an extension module plugged into its side slot, glowing green. Four single receipts lie in front of it, green threads connect each to its line on the card statement.

Jira apps, agent runs, thesis, games

More work

Ten Jira Cloud apps, all listed on the Marketplace

Sole developer and publisher 9 of 10 without write access at most 2 permissions each 2026

Ten Jira Cloud apps, a measurement of where coding-agent runs spend their time, my Bachelor’s thesis on CNN object detection, and two game prototypes.

A pinboard and bench with more work: thesis prints next to a microscope, game cards, and a wall clock whose face is a pie chart of where coding-agent time goes.

05 · About

About

My background is SAP analytics and integration: SAP BW/4HANA and SAP Analytics Cloud, data models, interfaces and custom code. I also run AI workshops for colleagues. Outside work, I build my own software with coding agents.

Role
SAP & AI Consultant · AI engineering · Heidelberg
At work
Member of the AI competence team and SAP AI ambassador. Internal AI assistants with SAP and company knowledge that colleagues use daily.
Study
B.Sc. Applied Computer Science, Heidelberg University. Bachelor’s thesis, graded 1.0
Stack
SAP BW/4HANA, SAP Analytics Cloud, TypeScript, Python, Atlassian Forge, coding agents
Languages
German (native), English (C1)

Career as a timeline

06 · Contact

Contact

Questions, comments, ideas or just a conversation: always glad to hear from people.

hallo@dangelstudio.de GitHub Back to the start

Legal notice · Privacy

The whole study from a distance in warm evening light: drawing board, tea and basket on the left, the work stations along the pegboard to the right.

That was the tour 06 / 06

Heidelberg in the evening: castle, old bridge, river.

How this was made

Built with coding agents: the 3D scene, the painted window view and the devices are code written with Claude Code; the printer and the SAP terminal also use details from OpenAI Codex and Kimi. Furniture and props are Poly Haven scans (CC0). Seal photos, sounds and music are CC0 or public domain. The KERBHOLZ card illustrations are AI-generated.

hallo@dangelstudio.de Plain site

Scroll on to start againBack to the start

Back to the start