Service

Applied AI/ML Pilot

A focused, system-independent pilot for computer vision, sensor and time-series analysis, knowledge retrieval, document extraction, classification, forecasting or decision support. The work starts with the real workflow and a baseline, then evaluates performance, failure modes, ownership and integration—not just whether a demo looks impressive.

Discuss an AI/ML pilot View services

Call when A specific AI/ML use case needs to be proven against real data and daily work.

What I do I define the decision and baseline, prepare representative cases, build or coordinate a focused prototype, evaluate results and document guardrails, limitations and the path into the existing workflow.

You receive Working pilot, representative test set, evaluation results, documented failure modes, guardrails, integration recommendation and a clear decision to harden, integrate or stop.

Your work package

What we agree before starting

Defined scope

One defined decision or workflow, with a baseline, representative evaluation set and agreed success and stop criteria.

How the result is accepted

Run the agreed evaluation against the baseline, expose difficult cases and failure modes, and document whether to continue, narrow the scope or stop. A negative result can fulfil the pilot.

Production and ongoing support

The result is an evaluated pilot, not an automatic production release. Hardening, live integration, operational monitoring and support are scoped after the evidence is reviewed.

Typical starting points

Concrete issues this work can handle

The service is useful when the issue is specific enough to inspect in real systems, reports, data samples or daily workflows.

Distinguish normal pump cycles from smaller leak patterns in water-pressure data

Labelled pressure histories establish what normal demand and pump starts look like. The pilot then measures whether smaller, persistent deviations can be detected early without creating an unusable false-alarm rate.

Capture hoof images in a barn passage and flag visible irregularities for human review

The pilot defines usable camera positions, image quality and visible review criteria before evaluating detection performance. Uncertain or safety-relevant findings remain with a qualified human reviewer.

Classify inbound requests and route them with review thresholds

Representative requests are used to measure routing quality by category, language and difficult edge cases. Low-confidence items go to a review queue instead of being sent automatically.

Extract structured fields from operational documents

A clear target schema and field-level evidence make extraction testable. Messy layouts, missing values and conflicting content are included so uncertain results can be flagged rather than silently accepted.

Build a controlled document assistant that returns source context

Retrieval is tested for source coverage, permissions, traceable context and the ability to abstain. The assistant should support a decision without presenting unsupported text as operational fact.

Test anomaly detection or forecasting against representative historical data

Historical data is split in time and compared with a simple baseline. Evaluation covers drift, false alarms and operational lead time before forecasting or anomaly detection is considered useful.

Specialist validating machine-learning results, operational data and document workflows

Applied AI & ML projects

From sensor and camera data to EAM and Excel

The examples range from pressure and camera data to EAM log analysis, text assistance, Excel and score-based scheduling. The two physical projects are shown in more detail.

Concept illustration of a pressure sensor, water pump, Arduino-style controller, Linux mini computer and two pressure curves — Concept illustration reconstructed from the original pressure-sensing prototype.

Time-series machine learning

Small-leak detection from water-pressure cycles

Normal water use lowered the line pressure until the pump restarted and restored it. The challenge was to distinguish that expected cycle from the slower, persistent pressure patterns associated with smaller leaks.

Build: Pressure sensor → Arduino → Linux mini computer
ML approach: Time-series features capturing water use, pump starts, pressure recovery and persistent drift.
What it demonstrated: In the prototype, the model separated ordinary draw-and-recovery cycles from deviations consistent with smaller leaks.

A model alert was intended to prompt an inspection, not replace verification of the pipework.

Self-initiated project · still active · now used in a slightly adapted form

Concept illustration of a horse walking past a low-mounted camera with a sequence of hoof images marked for review — Concept illustration reconstructed from the original camera-based hoof prototype.

Computer vision

Camera-based hoof anomaly screening

A low-mounted camera captured horses’ hooves as they passed through a barn walkway. The system analysed the captured hoof regions and surfaced visible irregularities or possible signs of injury for human review.

Build: Low-mounted camera → image capture sequence → local image processing
ML approach: Computer-vision detection of hoof regions and screening of visible features in the captured frames.
What it demonstrated: A routine barn passage could provide a repeatable screening point and surface frames worth checking.

Screening support only. A flagged image is not a veterinary diagnosis and always requires a qualified human assessment.

Watch the early hoof prototype

The video shows the early test setup before the camera was permanently installed.

Self-initiated project · still active · now used in a slightly adapted form

More applied AI

AI assistance inside EAM, Excel and planning

These projects show how AI and data-based recommendations can be integrated into existing tools without turning every use case into a separate platform.

EAM · AI-assisted log analysis

Automatic Attune EAM log analysis

Attune EAM log files are read automatically and relevant entries are presented as dashboard hints for further investigation.

EAM · text assistance

AI-assisted text improvement inside EAM

A lightweight AI assistant improves short descriptions and notes directly inside EAM and returns a clearer wording suggestion.

Excel · AI agent

AI agent inside Excel

The AI agent is integrated into an Excel task pane and provides assistance without leaving the workbook. It remains in use and is continuously improved.

In use · continuously improved

Watch the Excel AI agent demo

EAM · decision support

Automatic scheduling with calculated scores

For possible employee and task assignments, the system calculates scores from defined criteria and recommends a suitable match in HxGN EAM.

Watch the scheduling and scoring demo

Scope and deliverables

Scope and tangible results

The scope stays close to the systems, data and workflows that affect day-to-day work. The deliverables are designed to support a clear next decision.

Scope

Computer vision for visible anomalies and review workflows

Sensor and time-series models for patterns, drift and anomalies

Document extraction and classification

Controlled knowledge retrieval with source context

Evaluation on representative cases and human-in-the-loop guardrails

API and workflow integration

Deliverables

Working pilot tied to one real workflow

Baseline and explicit evaluation criteria

Results across representative and difficult cases

Documented limitations and failure modes

Human guardrails and ownership boundaries

Integration recommendation and clear next step

How it works

How the engagement works

A defined sequence keeps the work focused. Representative examples and access to the right owners make the review faster and more reliable.

Steps

Define the workflow, decision, baseline and stop criteria

Prepare representative data or documents and a test set

Build the smallest useful pilot

Evaluate normal, difficult and failure cases

Document guardrails, ownership and integration options

Decide whether to harden, integrate, reshape or stop

Fit and boundaries

Good fit and clear boundaries

A focused scope makes the work easier to evaluate, deliver and hand over.

Good fit

Operations, data or automation teams with one defined AI/ML use case
Organisations that need evidence before funding production work
Partners who need focused AI/ML delivery inside a wider programme

Not included

Not a chatbot added without a defined workflow
Not a model demo presented as a production solution
Not predictive maintenance without adequate history and labels
Not autonomous high-risk decisions without accountable review
Not model training for its own sake

Related services

Useful next steps and related services

Operational problems can cross system, data, interface, workflow and AI/ML boundaries. These services are often relevant together, but each can also stand alone.

Does this service fit your case?

Bring the current workflow or problem, a few representative examples and the business context. The first step is to decide what is worth changing, testing or leaving alone.