Posts

Within this test, the open-source framework unstructured is used to evaluate the extraction process of text from structured documents. The goal is to assess how suitable unstructured is for practical use in AI-based information systems – especially with respect to text extraction, semantic preparation (chunking/tokenization) and subsequent embedding generation for vector-based retrieval systems. Here is an example of a PDF file that was used for analysis with unstructured. pm-partnerschaft-stackitDownload To run the unstructured library, the official Docker image is used. It contains all required dependencies (e.g. Tesseract, Poppler, Python libraries) and allows immediate use without a local Python installation. ...

I started this project to learn React in a practical way – and not just follow tutorials. I wanted to implement a realistic frontend scenario that includes typical requirements like user authentication, data management, user interactions, and dynamic UI components. Instead of building my own backend, I consciously chose Supabase – a backend-as-a-service platform that is ideal for learning and prototyping purposes. This allowed me to fully focus on the React ecosystem, including routing, state, component structure, and responsive UI. ...

MindsDB is an open-source platform designed to enable machine learning, time series analysis, and the integration of large language models directly into traditional database workflows. The platform makes AI functionality accessible through simple SQL queries without requiring a separate machine learning infrastructure. In this post we introduce one of the many features of MindsDB: calling a large language model (GPT-4) via a predefined template that is dynamically populated with database values. The goal is to automatically assess the emotional impact of songs from an existing Spotify dataset. Only a portion of MindsDB’s full feature set is used to demonstrate the basic workflow and the interaction between the database and the LLM. ...

The transmission of facial expressions in real time to digital characters is an important component of modern animation and visualization processes. With Epic Games’ Live Link Face App and Unreal Engine 5, the facial movements of a real person can be precisely transferred to a digital Metahuman character. A prerequisite for this is an iPhone with an integrated TrueDepth camera that is connected via a local network to the computer running Unreal Engine. This tutorial shows how to set up the Live Link Face App and connect it to the engine, how to correctly prepare the Metahuman, and how to finally transmit the facial data live. The goal is to establish a working real-time connection in which the Metahuman moves synchronously with the facial expressions of the real person. ...

Audio2Face is an AI-powered tool within NVIDIA Omniverse specifically designed to generate realistic facial animations based solely on audio. It is part of the Omniverse platform, which provides a real-time collaboration and simulation environment for 3D workflows. Audio2Face uses a neural network to automatically convert spoken language into lively facial expressions and movements. Typically, Audio2Face is used to make characters in games, films, or digital avatars speak without complex keyframe animation. The generated movements can either be used directly or transferred to custom 3D characters, which is particularly interesting for virtual productions, digital twins, or interactive applications. ...

In this project, I developed an AI agent that automatically analyzes events from the NRW region and filters them according to personal criteria. The goal was to filter out only those events that are truly relevant – based on an individually defined prompt. This image shows the list of over 350 events taking place on a single day in Düsseldorf. Across all of NRW, there are several thousand events in one day. ...

The question was whether current Large Language Models (LLMs) such as GPT-4 or DeepSeek are able to automatically and reliably classify music tracks, specifically salsa songs, based on title, artist, lyrics, and metadata into „Salsa Cubana“ or „Salsa Línea“. It was known that the available information (metadata, genre tags, lyrics) is incomplete and partly inconsistent. The test was explicitly designed to determine the practical limits of today’s LLMs in this context. ...

What is Omniverse? Omniverse is a platform from NVIDIA that allows you to create, connect, and simulate virtual 3D worlds – all in real time. Omniverse is an open platform for developers, designers, engineers, researchers, and creatives to: Connect 3D applications (e.g. Blender, Maya, Unreal Engine) Collaborate in a single scene – live and simultaneously. Create physically realistic simulations and AI-driven applications. What is Omniverse used for? Design, visualization & simulation of objects such as vehicles in real time. Digital twins of factories for process optimization. Asset creation and visual simulations. Training with synthetic data, e.g. for autonomous vehicles. Real-time collaboration in animation and rendering. ...

ComfyUI is a node-based user interface for controlling and modifying AI models for image and video creation. Wan 2.1 is a text-to-video model (T2V) specifically developed for generating videos based on text inputs. This guide provides step-by-step instructions on how to set up ComfyUI with Wan 2.1 locally. Each section explains the required components, why they are necessary, and how to install them correctly. This guide assumes Python 10 and a GPU with CUDA support. ...

AI models often produce unstructured or imprecise answers. Anyone who wants better results must adjust their prompts accordingly. One way to do this efficiently is with prompt decorators – clear instructions at the beginning of a prompt that control the AI’s response behavior. In this post, I show how to teach the AI to understand these decorators and how to use them afterward. Explain Prompt Decorators to the AI The AI is given a clear definition of the decorators, for example „+++StructuredAnswer“, so that it understands their meaning. The instruction to apply them in future answers ensures that they don’t apply to just a single question. If the AI doesn’t have long-term memory, this introduction must be repeated in each new session. ...

In this project, I developed a Thunderbird extension that uses ChatGPT for advanced spam detection. Incoming emails are automatically analyzed and classified according to various criteria. A local Flask server handles the communication with ChatGPT and assesses whether a message should be classified as spam. The implementation serves as a demo to explore the possibilities of AI-powered filtering in Thunderbird. Workflow As soon as Thunderbird receives a new email, the extension becomes active. The message is intercepted before it is viewed by the user. The extension extracts the subject, the sender, and the email body. ...

Here Ollama was run with NVIDIA-GPU-support under Docker on a Windows-11-system. OpenWebUI was used as a user-friendly interface to operate AI models locally. OpenWebUI offers the advantage that users can easily switch between different models, manage requests and conveniently control AI usage through a graphical interface. It also provides a better overview of running instances and facilitates testing different models without manual configuration changes. Install WSL 2 Install NVIDIA CUDA Drivers In order for Docker containers to access the GPU, the NVIDIA Container Runtime is required. This enables faster and more efficient computation of AI models, since compute-intensive processes are handled not by the CPU, but by the more powerful GPU. https://developer.nvidia.com/cuda/wsl ...

This code shows how an artificial neural network is trained with the MNIST dataset to classify handwritten digits (0-9). The goal is for the model to be able to predict which digit is shown based on the image data. This is achieved by: 1. Loading and preprocessing of the MNIST image data. 2. Creating a neural network with multiple layers (Layers). 3. Training the network with training data. 4. Evaluating the model’s performance on test data. 5. Testing the model on new sample data. ...

Download the Ollama LLM runtime environment download and install it. After installation the server can be accessed at http://127.0.0.1:11434/. 3. Show the list of installed models. The list should be empty. ollama list 4. Download the llama3.2 LLM and DeepSeekv3 (404 GB HD & 413 GB RAM). ollama pull llama3.2 ollama pull deepseek-v3 On the Meta website you can find the current versions of the LLM. 5. Start llama3. ...

Send a question to OpenAI via Spring AI and display the answer Create OpenAI key https://platform.openai.com/settings/organization/api-keys Then set the key as an environment variable: OPENAI_API_KEY Create a new Spring Boot project: https://start.spring.io/ Within the Spring Boot application, i.e., in the “application.properties” file, reference the OpenAI key, i.e., environment variable (OPENAI_API_KEY). After creating the interface and classes, the project structure should look as follows: After running the unit test, the answer to the question “Who would win in a fight between Superman and Chuck Norris?” should be displayed. In this case: ...

In this post, I explain how you can use Whisper, an AI-based tool from OpenAI, for automatic transcription of videos. Whisper is capable of accurately converting spoken language in various languages – including German – into text. This makes it ideal for transcribing, for example, interviews, lectures, or personal videos. Install Python 3.10 Whisper requires the Python programming language and needs a version between 3.7 and 3.10. In this guide, we use Python 3.10 to avoid compatibility issues. ...

Access advanced Firefox settings: about:config Set network.trr.mode from 0 to 2. Set network.trr.uri to https://mozilla.cloudflare-dns.com/dns-query. network.trr.mode is a configuration setting in Firefox that controls the use of DNS over HTTPS (DoH). TRR stands for Trusted Recursive Resolver and refers to the use of DoH to send DNS queries over an encrypted HTTPS connection instead of traditional, unencrypted DNS queries. 0 – DoH is disabled: Firefox uses only traditional, unencrypted DNS (over UDP or TCP) and sends no DNS queries over HTTPS. ...

This blog post covers the concept of embeddings and vector databases. Initially, it explains what embeddings are and how they are used in the field of Natural Language Processing (NLP). Then, an explanation of vectors in a three-coordinate space and their extension to multidimensional vectors follows. Finally, ChromaDB is introduced, a specialized vector database. What is an Embedding? An Embedding is a technique in the field of machine learning and data processing that aims to transform objects such as words, sentences, or documents into a continuous vector space. In this vector space, similar objects are represented by similar vectors, meaning they lie close together. Embeddings are frequently used to capture and analyze the semantic meaning of texts. ...

Forward requests 1:1 to ChatGPT via Streamlit and display the response. streamlit run ChatUI.py

Send Message via LangChain to ChatGPT Here we send a message via LangChain to ChatGPT and output the response. Perform Chaining with “chain” Use “chain” to link the model and the parser together. Template Prompts Prepare the prompts using a template. Use the template: The target language is externalized: Further Chaining Chain the template with model & parser: Server via FastAPI & REST Endpoint Create a server via FastAPI. The server loads the application (app.py), handles the REST request and outputs the translation: ...