Use Cases
A curated set of AI based systems. Each case highlights context, constraints, technical decisions, and lessons learned.
On this page
AI Based Meter Reading
Manual meter reading is time consuming and error prone at scale. The objective was to design an AI based system that automates meter reading from images while maintaining accuracy, transparency, and human validation.
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
The solution was conceptualized as an end to end pipeline. A custom object detection model was trained from scratch to locate relevant fields. A separate OCR model was fine tuned on custom data. Labeling, post processing, and manual validation were designed to create a feedback loop for continuous improvement.
Outcome and reflection
The resulting system extracted structured meter information across varying conditions. Combining detection with fine tuned OCR improved robustness compared to generic solutions. The main learning was that labeling strategy and validation workflows are as important as model architecture for production quality.
AI Based Chip Detection System
In a manufacturing environment, unrecognized chip accumulation during turning processes can lead to quality issues, downtime, and increased manual inspection. The goal was to design an AI based system capable of detecting and classifying chip related anomalies directly in production conditions.
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, starting with data acquisition and labeling in real conditions. A computer vision pipeline was designed with a focus on robustness. Beyond model training, the solution included a Linux based inference server, an operator interface, and a mobile data acquisition setup to continuously improve training data.
Outcome and reflection
The resulting system achieved reliable chip detection and classification in operational conditions. It provided real time feedback to operators and established a foundation for scaling similar inspection solutions. The key learning was that model performance, data quality, hardware constraints, and user interaction all influence the final outcome and should be treated together.
RAG Chatbot with Custom Fine Tuned Models
Accessing structured and unstructured knowledge across systems is often fragmented. The objective was to build a retrieval augmented generation system for contextual question answering over domain knowledge, with control over data usage, model behavior, and system transparency.
Constraints
- Heterogeneous data sources including text and graph structured data
- Domain specific semantic understanding
- Traceable and auditable responses
- Interactive performance requirements
Approach
The system was designed as a modular RAG architecture combining retrieval, embedding similarity search, and controlled generation. Open source embedding models were fine tuned for domain retrieval. Generative models were fine tuned to align responses with domain terminology. A graph database enriched retrieval using explicit relationships.
Outcome and reflection
The chatbot produced coherent responses grounded in the underlying sources. Domain relevance improved compared to generic embeddings and base models. The main learning was that retrieval quality and structured context often influence answer quality more than model size.
Time Series Anomaly Detection on Edge Microcontroller
Early anomaly detection in sensor signals is critical for reliability and preventive maintenance. The goal was to build a deep learning based anomaly detection system for time series data and deploy it on a resource constrained microcontroller.
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.
Outcome and reflection
The deployed system detected anomalous patterns in real time under constrained hardware conditions. Edge inference reduced latency and avoided transmitting raw sensor data. The key learning was that aligning model complexity with deployment reality often matters more than maximizing accuracy in isolation.