Applied AI consulting: RAG, agents, evaluation, edge AI and plant vision

We build AI systems that have to be right about something physical, and we prove whether they are. Two developers, everything below built and run in-house for our own products first.

WHAT WE DO

Applied AI systems. Retrieval over your own business data, tool-use and MCP layers, agent orchestration, and model routing that decides when a cheap model answers and when an expensive one has to. We run a routed eleven-model stack in production.

Evaluation and validation. Where most AI projects have nothing. Eval suites, gold sets and holdouts that do not leak, LLM-judge ensembles for labelling and edge-case audit, prompt versioning and tracing, calibration measurement, and abstention thresholds tied to what a wrong answer costs.

Edge and IoT. Models small enough to run on a sensor node rather than in a data centre: fine-tuning and LoRA, quantisation, GGUF and ONNX export, on-device inference. We build the hardware side too, sensor hubs and controllers, so the model and the device are designed against each other.

Vision for plants and biology. Dataset construction, annotation pipelines, classifier training, confusion analysis, deployment. Our own corpus is 259,000 labelled images across 37 classes from 15 sources.

Community platforms with AI. Moderation and safety layers, ranking, and the parts that have to work unattended at scale.

WHAT WE ARE KNOWN FOR Directional error. When one symptom has several causes, a share of model errors are not near misses but reversals: the system tells a grower to feed more when the plant needs less. No benchmark reports this. We name it, measure it on your data, and control it with directed rules and calibrated abstention. Standalone audit or part of a build.

EVIDENCE Production vision pipeline serving real users. Held-out accuracy 94.3 percent on the 4-class trunk, full confusion matrix on file. Calibration error 0.113, temperature fitted on a 644 case holdout the model never trained on. Around 300 tracked training runs since 2025. 10,000 H100 GPU hours granted by the IT Center of RWTH Aachen University in September 2026 after technical review, with two further EuroHPC applications under review.

HOW WE WORK Fixed fee per engagement, scoped after a short call. Your model, your data and your results stay yours. We publish nothing without written consent. Remote, English or German, EU based.

GOOD FIT FOR Companies shipping an AI feature that gives advice someone acts on. Teams that need to defend a model to a customer, a partner or an auditor. Agtech, greenhouse technology, crop protection, connected devices, and any domain where the ground truth is biological rather than clean.