Location Matters
Case Study: libgeo2pixel
The libgeo2pixel library implements all required reference frames, coordinate systems, and the three geolocation algorithms (Ray Marching, Ray Intersection, Gradient Descent). It is written in C, optimized for speed and easy to port. Rotations are performed only with quaternions, making it suitable for real‑time use. Its only dependencies are the C standard library and a small interface to the gdalinfo tool for accessing DEM (digital elevation model) data; otherwise it is completely self‑contained. This allows the library to run on smartphones, drones, or other consumer devices.
CamTraps 2024 Paper
Case Study: Perceptomat
The video stream of a drone is analyzed in real time by AI-powered software. Prompt-based object detection is used to define potential indicators (e.g., specific objects, traces or patterns). Identified indicators are geo-located and related to an upstream causal risk model (Riskomat) to detect complex scenarios, e.g. when indicators are harmless on their own but the threat increases in spatial combination. The prototype operates exclusively with off-the-shelf hardware and does not require expensive special sensors. Application areas can be found in emergency services organizations, for example.
Präsentation: Geolokalisierung von Objekten in Drohnenvideos (IKT Sicherheitskonferenz 2025)
Case Study: Gridomat (ongoing)
Gridomat is a self‑contained, web‑based platform that processes video streams from multiple drones simultaneously. Unlike Perceptomat’s single‑drone system, the platform aggregates data in real time, applies the same prompt‑driven object detection, and gathers evidence for the causal risk model (Riskomat). Parallel analysis creates a more comprehensive situational picture, enabling rapid, evidence‑based risk assessment and evacuation decisions. The solution runs entirely in the browser, requires no special infrastructure, and scales flexibly with the number of connected drones.
Case Study: jlens-agent
The jlens-agent prototype is a concept study on applying Jacobian Lens (Anthropic, July 2026) as an introspection method for a large language model (LLM). Combined with a causal model in Riskomat, we use it to detect, interrupt and, if necessary, escalate the intentions of an AI agent before they are executed. From a given policy, two sensors are generated that process Jacobian-Lens data and raise alarms. The semantic analysis of alarms is performed by a second LLM. In case of a detection, the alarm is propagated to the causal risk and decision model in Riskomat.
Präsentation: Präkognition für KI-Agenten: Lassen sich gefährliche Absichten erkennen und stoppen? YouTube: Demo jlens-agent (IKT 2026)