Self-Training Large Language Models for Tool-Use Without Demonstrations
We propose a self-training approach that enables large language models to learn tool use without requiring human demonstrations. Our method uses self-…
How Real-Time Quality Inspection with AI Is Transforming Manufacturing
Product quality is one of the most critical drivers of cost, customer satisfaction, and brand reputation in manufacturing. Yet traditional inspection methods—whether manual or rules-based automation—struggle to meet the scale, speed, and accuracy demanded by today’s production environments. Real-time quality inspection using artificial intelligence and computer vision is changing the landscape. Miniml helps manufacturers deploy intelligent visual systems that inspect every product on the line with precision. By applying deep learning models trained on real production data, we enable fast, scalable, and highly accurate defect detection—reducing waste, increasing throughput, and improving consistency across operations. Why Manual Inspection Doesn’t Scale Manual inspection is time-consuming, inconsistent, and expensive. Human inspectors are subject to fatigue, distraction, and variability in judgment. Even well-trained operators can miss subtle or rare defects, especially on high-speed lines or in variable lighting conditions. Basic rule-based vision systems—often built on fixed thresholds or shape-matching algorithms—offer limited flexibility and poor generalization to new defect types, product variations, or environmental changes. This is especially problematic in sectors like: In these industries, a missed defect can mean millions in recalls or regulatory penalties. That’s why manufacturers are turning to AI. How Real-Time AI Inspection Works AI inspection systems built by Miniml use convolutional neural networks (CNNs) and other deep learning architectures to identify visual anomalies in real time. These models are trained on large sets of labeled production images, allowing them to learn what “good” and “bad” look like without relying on hard-coded rules. Key features of our systems include: Models can be trained for binary classification (pass/fail), multi-class labeling (defect types), or segmentation (precise defect outlines), depending on the use case and required precision. Common Applications of AI Visual Inspection AI-based inspection can be adapted to nearly any visual quality control task, including: We deploy these systems at various stages in the production process—on packaging lines, inline with assembly equipment, or at end-of-line QA stations. What Makes a Good AI Inspection Model? Performance depends on more than model architecture. The success of a real-time inspection system also hinges on: We work closely with plant teams during data collection and model training to ensure real-world reliability—not just lab performance. Deployment: From Edge to Cloud Miniml supports deployment options tailored to your IT and operational requirements. Many of our clients run real-time inference on industrial edge devices (e.g., NVIDIA Jetson, Intel Movidius) directly on the line. Others opt for centralized cloud inference with APIs integrated into SCADA or MES systems. We offer features like: Security, latency, and maintainability are factored into every deployment plan—whether you’re running a pilot or global rollout. Buying vs. Building: Why Custom Models Matter Generic inspection systems often fall short in specialized environments. Pretrained models may not understand your unique products, packaging materials, or acceptable tolerances. Miniml builds custom-trained models that adapt to your actual operating conditions—your lighting, your equipment, your edge cases. That’s how we achieve enterprise-grade performance, not just benchmarks. Integration Considerations for Plant Teams We work directly with manufacturing, automation, and quality teams during planning and deployment. Key integration factors we support include: Our systems are designed for uptime and simplicity, with fallback states and diagnostics built in. Frequently Asked Questions Can AI models detect new defect types?Yes. We use anomaly detection models that flag “unseen” issues even if they weren’t in the training set—perfect for continuous learning environments. What happens if lighting conditions change?Our models are trained on a range of lighting conditions, and we support dynamic calibration routines to maintain accuracy over time. What if I don’t have labeled data?We help you collect and label high-quality datasets from your line. We can also accelerate this process using weak supervision and human-in-the-loop review. Start Building Smart Quality Systems AI-powered quality inspection is no longer experimental—it’s a proven, scalable solution delivering measurable impact across industries. Whether you’re inspecting food packaging, automotive parts, or PCB assemblies, Miniml can help you move from manual checks to continuous, intelligent QA. Book a Consultation
When Can Proxies Improve the Sample Complexity of Preference Learning?
We provide theoretical and empirical analysis of when proxy rewards can improve sample efficiency in preference learning. Our work establishes conditi…
Is Complex Query Answering Really Complex?
We challenge conventional wisdom about complex query answering by demonstrating that many supposedly complex queries can be solved with surprisingly s…
An Auditing Test to Detect Behavioral Shift in Language Models
We develop a comprehensive auditing framework to detect behavioral shifts in language models across different contexts and time periods. Our approach …
TuBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning
We investigate the cross-lingual transferability of backdoor attacks in instruction-tuned large language models. Our findings reveal that backdoors ca…
Steering Knowledge Selection Behaviours in LLMs
Read Paper Knowledge Selection in LLMs We investigate how large language models (LLMs) select and utilise knowledge when generating responses. Our analysis reveals that LLMs exhibit systematic biases in knowledge selection, often favouring certain types of information over others regardless of relevance or accuracy. Through controlled experiments using knowledge-steering techniques, we demonstrate that it’s possible to influence LLMs’ knowledge selection behaviours. We introduce novel methods for steering models towards more balanced and contextually appropriate knowledge utilisation, significantly improving response quality and factual accuracy. Our findings have important implications for developing more reliable and controllable language models, particularly in knowledge-intensive applications where accurate information retrieval and utilisation are critical.
Mixtures of In-Context Learners
We introduce Mixtures of In-Context Learners (MiCL), a novel approach that combines multiple in-context learning strategies to improve few-shot perfor…
Low-rank lottery tickets
Low-rank lottery tickets: finding efficient low-rank neural networks via matrix differential equations Read Paper Neural networks deliver exceptional performance but can be impractical for applications with limited hardware or energy resources due to their high memory and computational demands. This work introduces a novel algorithm to identify efficient low-rank subnetworks during the training phase, significantly reducing both training and evaluation costs. The approach restricts weight matrices to a low-rank manifold and updates only the low-rank factors during training. Using dynamic model order reduction techniques, the method ensures approximation, stability, and descent guarantees. It also adapts ranks dynamically throughout training to maintain the desired accuracy. Numerical experiments on fully-connected and convolutional networks demonstrate the efficiency of this technique