Context
Manual meter reading is time consuming and error prone at scale. Images vary strongly across meter types, lighting conditions, layouts, and image quality.
Selected applied AI cases
Work
A curated set of AI based systems. Each case highlights context, value questions, constraints, technical decisions, operational considerations, and lessons learned.
Value question
Can relevant meter information be extracted reliably enough to reduce manual effort while keeping validation and correction paths clear?
Constraints
- High variability in meter types, layouts, and visual conditions
- Extraction of meter number, meter reading, OBIS, and barcodes
- Limited labeled data at project start
- Need for human-in-the-loop validation
Approach
A custom object detection model was trained to locate relevant fields. A separate OCR model was fine tuned on custom data. Labeling, post processing, and manual validation were designed as a feedback loop for continuous improvement.
Operational considerations
The system needed more than accurate predictions. It required traceable outputs, structured correction, validation workflows, and a process for turning errors into better training data.
Outcome and reflection
The resulting system extracted structured meter information across varying conditions. The main learning was that labeling strategy and validation workflows are as important as model architecture for production quality.
Context
In turning processes, unrecognized chip accumulation can lead to quality issues, downtime, and increased manual inspection effort.
Value question
Can visual detection provide reliable real-time feedback for operators and create a basis for scalable inspection solutions?
Constraints
- Limited and noisy visual data from industrial environments
- Real-time inference requirements on edge hardware
- Integration into existing production and monitoring systems
- Clear feedback for machine operators
Approach
The system was developed end to end, from real-condition data acquisition and labeling to a robust computer vision pipeline. The solution included a Linux based inference server, an operator interface, and a mobile data acquisition setup.
Operational considerations
Model behavior, hardware constraints, data quality, operator interaction, and integration with production processes had to be treated together rather than as separate concerns.
Outcome and reflection
The system achieved reliable chip detection and classification in operational conditions. The key learning was that production AI succeeds only when the model, data loop, deployment environment, and user feedback are aligned.
Context
Accessing structured and unstructured knowledge across systems is often fragmented, especially when domain terminology and relationships matter.
Value question
Can a retrieval augmented architecture provide useful answers while keeping source grounding, transparency, and domain relevance under control?
Constraints
- Heterogeneous data sources including text and graph structured data
- Domain specific semantic understanding
- Traceable and auditable responses
- Interactive performance requirements
Approach
A modular RAG architecture combined retrieval, embedding similarity search, graph enriched context, and controlled generation. Open source embedding models were fine tuned for domain retrieval, and generative models were aligned with domain terminology.
Operational considerations
Retrieval quality, confidence thresholds, source traceability, and validation patterns were treated as core design decisions rather than secondary features.
Outcome and reflection
The chatbot produced coherent responses grounded in underlying sources. The main learning was that retrieval quality and structured context often influence answer quality more than model size.
Context
Early anomaly detection in sensor signals is critical for reliability and preventive maintenance, but raw signals are noisy and failure modes are rare.
Value question
Can anomalous patterns be detected in real time on constrained hardware without relying on continuous raw data transmission?
Constraints
- Noisy and highly variable time series data
- Limited labeled examples for rare failure modes
- Strict memory and compute limits on microcontroller hardware
- Real-time inference and low power requirements
Approach
Data exploration and preprocessing established a stable training dataset. Multiple architectures were evaluated with attention to robustness and model size. The chosen model was optimized and deployed using a lightweight runtime for on-device inference.
Operational considerations
The model had to fit hardware limits, latency requirements, power constraints, and maintenance expectations. Accuracy alone was not a sufficient optimization target.
Outcome and reflection
The deployed system detected anomalous patterns in real time under constrained hardware conditions. The key learning was that model complexity must be aligned with deployment reality.