We connect software engineering, AI research and interactive learning.

Our work spans SpatialAgents and GIS applications, the evaluation of AI agents and interactive learning paths for numerical methods.

Discuss a project Explore our fields

SpatialAgents Benchmark 08 · climate and hazard analysis in QGIS · risk and evacuation layers from a benchmark run

Flagship 01 · SpatialAgents

SpatialAgents work where the geodata lives.

SpatialAgents connect QGIS Desktop to Claude Code, Codex CLI and OpenCode through a technical bridge.

Each supported agent system can launch multiple SpatialAgents to work on geospatial tasks in the shared workspace. The agents gain a controlled route to project state and geospatial tools. We develop the Geospatial API for AI agents as a library of its own; the SpatialAgents import it directly in the shared workspace. This lets agent systems apply their native code-generation and verification abilities without an additional MCP server and with low token overhead. Its self-describing, fault-tolerant interface supports autonomous correction through specific errors and suggested alternatives. Feedback from practical agent analyses flows continuously back into the API, tests and agent skills.

We research and develop. Projects, setup, training and support are provided by Geoinformatikbüro Dassau GmbH (GBD).

agent workspaces

Explore SpatialAgents in detail

Codex shell in the SpatialAgents workspace showing available QGIS and geospatial tools — SpatialAgents / CodexCodex and the agents it launches use QGIS and geospatial tools through SpatialAgents in the shared workspace.

Flagship 02 · CAD-Agents

Evaluate CAD reconstructions generated by AI agents across five dimensions.

The research testbed measures dimensional accuracy, feature fidelity, shape fidelity, drawing fidelity and thread fidelity.

These six Claude Code runs with opus-5 achieved a shape fidelity score of 1.00. Every model shown was reconstructed from a technical drawing.

Claude Code reconstruction of a curved cover with a stepped central bore from case 045 — **Case 045**Claude Code · opus-5 · run 1 · shape fidelity 1.00

Claude Code reconstruction of a circular plate with a central bore and six cylindrical pins from case 014 — **Case 014**Claude Code · opus-5 · run 1 · shape fidelity 1.00

Claude Code reconstruction of a branched support with three bearing points and a reinforcing rib from case 027 — **Case 027**Claude Code · opus-5 · run 1 · shape fidelity 1.00

Claude Code reconstruction of a circular housing with a central tube and radial ribs from case 033 — **Case 033**Claude Code · opus-5 · run 1 · shape fidelity 1.00

Claude Code reconstruction of a ribbed bearing housing with four mounting points from case 080 — **Case 080**Claude Code · opus-5 · run 1 · shape fidelity 1.00

Claude Code reconstruction of a bearing block with a base plate, rib and transverse bore from case 900 — **Case 900 · bearing block**Claude Code · opus-5 · run 1 · shape fidelity 1.00

Select an image to open its high-resolution detail view.

Open the complete CAD-Agents site

Our fields

Development, research and learning.

Three fields connect concrete systems, methodical investigation and accessible technical education.

01 / Development

Expert software and AI systems.

A selection from our development portfolio: SpatialAgents, the Geospatial API for AI agents, Climate Risk and Vulnerability Assessment (CRVA) built on that framework, and Düngewende.

Explore development ↗

02 / Research

Study agents inside technical tools.

A selection from our research programme: SpatialAgents benchmarks, CAD-Agents and SpatialApp study the execution, evaluation and deep integration of AI agents in technical software.

Explore research ↗

03 / Learning

Make technical methods understandable.

The Learning Lab currently brings together three German-language interactive books on FEM, electromagnetism and bone remodelling.

Open the Holistech Learning Lab (German) ↗

Holistech Labs

What we measure, we measure ourselves.

Our measurement lab tests hardware, inference engines and models under controlled conditions and publishes the run logs.

The first report compares llama.cpp and halogen-flash-server running Qwen3.8-Flash-Next on an HP ZBook Ultra G1a: the same success rate across 20 tasks, but 486 against 157 minutes over the whole run.

Open Holistech Labs