AI Technologies & Tools.

OpenAI is the company that has decisively shaped the modern AI era. Founded in 2015 as a nonprofit organization and later converted into a for-profit structure, OpenAI develops some of the most capable large language models in the world — including the GPT series, the o reasoning models and DALL·E.

The current GPT models are particularly strong in code generation, multimodal tasks (text + image) and complex reasoning. Through the OpenAI API, these models can be embedded directly into your own applications, chatbots or workflows — controlled, scalable and with your own branding.

For professional integrations, vona uses the OpenAI API as one of its core technology components: precise, well documented, reliable.

Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei and other former OpenAI employees — with a clear focus on AI safety and the responsible development of AI systems.

The company develops the Claude model series and, with Constitutional AI, has established its own approach to making language models helpful, harmless and honest. Unlike many competitors, Anthropic regularly publishes safety research and takes an active part in the AI policy debate.

Anthropic is considered one of the few providers for which safety is not marketing but actively pursued research.

Claude is Anthropic's family of large language models — and one of the most powerful on the market. Claude stands out for its very large context windows (hundreds of thousands of tokens and more), nuanced, differentiated writing and particular strength in complex analyses and long documents.

The current models (Claude Opus, Sonnet, Haiku) offer a broad performance spectrum: from fast, inexpensive answers to deep, multi-step reasoning. Claude Opus is among the most powerful models available for complex tasks.

vona is happy to use Claude, especially in projects that call for nuanced writing, long contexts or especially precise reasoning.

Gemini is Google's current generation of multimodal large language models, developed by Google DeepMind. The model series — Ultra, Pro, Flash and Nano — covers the entire performance spectrum: from the lightweight on-device model to the most powerful variant for complex tasks.

Gemini's particular strengths lie in its deep integration with the Google ecosystem (Search, Workspace, Android), in multimodal capabilities (text, image, audio, video) and in real-time web access. Google also offers Gemini as a premium AI assistant for consumers and businesses.

For projects built on Google infrastructure or with multimodal processing requirements, Gemini is a natural choice.

Meta AI is Meta's AI assistant, integrated directly into WhatsApp, Instagram, Facebook and Messenger — making it one of the most widely used AI interfaces in the world. Powered by the Llama models, Meta AI can answer questions, generate images and help with everyday tasks — without leaving the familiar app.

Compared with standalone AI assistants, Meta AI scores with its ubiquity: billions of users reach AI capabilities right where they already communicate every day. For businesses, this opens up new possibilities in the WhatsApp Business context.

Google AI Studio is the official developer platform for the Gemini API — freely accessible to developers and businesses that want to integrate Gemini models into their own applications.

AI Studio lets you test prompts, adjust model parameters, try out multimodal inputs (text, image, audio) and generate code for the API integration directly. The platform offers a generous free tier for prototyping and smaller production projects.

For teams that want to experiment with Gemini quickly or build a proof of concept, Google AI Studio is the ideal starting point.

Meta Platforms (formerly Facebook) is one of the largest technology groups in the world — and at the same time one of the most significant open-source AI players. With the Llama model series (including Llama 2 and Llama 3), Meta has made powerful large language models freely available that can be run on your own hardware.

This has considerably advanced the democratization of AI: companies can run Llama models locally, fine-tune them and build them into their own products — without license fees and without transferring data externally. The larger Llama models are among the most powerful freely available models and can keep up with commercial models.

Mistral AI is a French AI company founded in 2023 that positions itself as a European alternative to OpenAI and Google. What sets Mistral apart: exceptional efficiency combined with high performance. At its release, Mistral 7B was the most powerful model in its size class.

Mixtral (Mixture of Experts) extended this approach: by selectively activating parts of the model, Mixtral achieves the quality of significantly larger models with significantly lower compute requirements. Many Mistral models are open source and can be run locally via Ollama.

For European companies with data protection requirements, Mistral is an attractive option — GDPR-friendly and operable on EU infrastructure.

Gemma is Google's family of lightweight open-source models — from a few hundred million to around 27 billion parameters — developed with the same technologies and training data as the Gemini models, but freely available and able to run locally.

Gemma models are optimized for local use and fine-tuning: compact enough for consumer hardware, yet powerful enough for many enterprise applications. Gemma models in some cases outperform significantly larger models in benchmarks and lend themselves to specialized adaptations.

With Ollama, Gemma models can be up and running locally in minutes — ideal for privacy-critical applications where no data may be transferred to the cloud.

Grok is the large language model from xAI, Elon Musk's AI company. What sets Grok apart from other models: direct real-time access to X (formerly Twitter) — allowing Grok to draw on current news, trends and posts in its answers, with no training data cutoff.

The current Grok models compete in benchmarks with the top models from OpenAI and Anthropic. Grok is integrated into X Premium and accessible via the xAI API. Earlier Grok versions have also been released as open-weights models.

For use cases where current social media data or real-time news is relevant, Grok offers unique possibilities.

DeepSeek is a Chinese AI research lab that caused a global stir in early 2025 with DeepSeek-R1: the model achieved performance on par with leading models — developed at a fraction of the usual training costs. This shook the assumption that leading AI models necessarily require hundreds of millions of dollars in training costs.

DeepSeek models (V3, R1) are freely available as open-weights models and can be run locally via Ollama. Particularly noteworthy: DeepSeek-R1 explicitly shows its thought process (chain of thought), which is valuable for traceable decisions and debugging.

DeepSeek's success has prompted the entire AI industry to re-examine its cost structure and training approaches.

Perplexity is an AI-powered search engine that combines real-time web search with large language models. Instead of a list of links, Perplexity delivers direct, summarized answers — always with source citations and linked evidence.

What makes it special: Perplexity uses multiple models (its own as well as OpenAI, Anthropic, Meta) and actively searches the web for every query. That makes it particularly strong for research, current information and fact-based answers.

In the GEO context (Generative Engine Optimization), Perplexity is one of the most important platforms: websites that Perplexity cites as trustworthy sources gain considerably in AI visibility. vona explicitly takes Perplexity into account in GEO strategies for clients.

Microsoft Copilot is Microsoft's AI assistant brand — powered by OpenAI models (GPT series) and deeply integrated into the Microsoft ecosystem. Depending on the context, this refers to different products:

Copilot in Microsoft 365 (Word, Excel, Outlook, Teams) helps with writing, analyzing and summarizing right in the flow of work. GitHub Copilot is the de facto standard for AI-assisted code completion in software development. Copilot in Windows is the system-wide AI assistant for end users.

For companies already invested in Microsoft 365, Copilot offers the lowest-friction path to AI adoption — the AI comes to the employees, not the other way around.

NotebookLM is Google's AI-powered notebook and document analysis tool — with a unique approach: instead of relying on general world knowledge, NotebookLM works exclusively with the documents the user uploads. This eliminates hallucinations stemming from training and delivers precise, substantiated answers.

Documents are analyzed, summarized, linked to one another and explained on request. The Audio Overview feature turns documents into a two-host podcast dialogue — a unique feature that puts complex content into an easily digestible form.

For knowledge management, research and the structured analysis of large volumes of documents, NotebookLM is one of the most practical AI tools on the market.

Nano Banana is the name of Google’s AI image model from the Gemini family. It creates and edits images based on text and reference images — from high-quality product images to motif variants and targeted image editing. Nano Banana Pro is the more powerful variant for higher quality and more complex instructions.

Areas of application: product images for online stores and social media, image variants for campaigns, model and scene visualizations. Among other things, vona uses Nano Banana in automated image pipelines for fashion online stores and in the desktop app RENEKI — with prompt building blocks, reference images and central cost overview.

Access is via the Gemini API, for example through Google AI Studio. Image rights, labeling and deliberate selection of the results are part of how we use it.

Ollama is the tool that radically simplifies running open-source large language models locally. With a single terminal command (ollama run llama3), models such as Llama 3, Mistral, Gemma, DeepSeek or Phi run on your own hardware — no cloud, no API costs, no external data transfers.

Ollama also provides a local REST API that mimics OpenAI's API format — enabling seamless compatibility with existing tools and integrations. n8n, Open WebUI and other tools can use Ollama as a local LLM backend.

vona recommends Ollama in three scenarios in particular: (1) privacy-critical environments (GDPR, medical data, legal advice), (2) high request volumes with high cloud API costs, (3) offline operation without an internet connection.

Hugging Face is the central platform of the open-source AI community — often described as "the GitHub of AI". The Model Hub hosts well over a million pretrained models for every conceivable task: language models, image generation, audio, video, classification and more.

Hugging Face's Transformers library is the most widely used Python library for accessing pretrained models and is considered the standard interface in AI research. Datasets offers thousands of public training datasets; Spaces enables the rapid deployment of AI demos and applications.

vona uses Hugging Face as its primary source for open-source models: when a client has a specific AI requirement, model selection often begins on Hugging Face — to check which pretrained model fits best before a custom solution is developed.

LM Studio is a desktop app that lets you download, try out and chat with language models on your own computer — without a terminal and without the cloud.

For developers it offers a local server with an OpenAI-compatible interface, so open models can be plugged into existing applications. The app runs on Mac, Windows and Linux; how fast the models respond depends on your hardware.

Open WebUI is a self-hosted chat interface in the style of well-known AI assistants. It can be connected to local models (for example via Ollama) and to cloud services.

Teams get their own privacy-friendly access to AI — with user management, conversation history and the option to bring in their own documents (RAG).

AnythingLLM is an application that lets you chat with your own documents: files, web pages and notes are prepared and provided to a language model as a knowledge source (RAG).

It runs as a desktop app or on your own server and works with local and cloud models. For small knowledge bases it is a quick start without having to build your own pipeline.

The Text Generation WebUI, usually called Oobabooga after its developer, is an open-source web interface for running and trying out many language models on your own hardware.

It supports various model formats and offers many settings. That makes it popular with tinkerers — for getting started, Ollama or LM Studio are usually simpler.

OpenRouter is a unified API layer through which more than 200 different LLMs from various providers (OpenAI, Anthropic, Google, Meta, Mistral and many more) can be reached under a single interface.

The practical benefit: model switching without code changes, automatic routing to the cheapest available model, resilience when providers go down and easy A/B comparison of different models. OpenRouter follows the OpenAI API format, which simplifies migrating existing integrations.

vona uses OpenRouter in projects that require model flexibility or where cost optimization through intelligent routing is important.

Snowflake is one of the leading cloud data platforms for data warehousing, analytics and data engineering. The architecture strictly separates compute and storage — allowing elastic scaling and multi-cloud operation (AWS, Azure, GCP) with a single SQL interface.

With Snowflake Cortex, the platform has integrated native AI capabilities: LLM functions directly in SQL queries, RAG pipelines on Snowflake data, vector search and the ability to deploy your own ML models. This enables AI-assisted data analysis without transferring data to external systems.

In addition, with Arctic, Snowflake has released its own open-source LLM, optimized for SQL and data analysis tasks.

Eden AI bundles services from many AI providers behind one uniform interface — not only language models but also image analysis, translation, speech recognition and more.

This makes it possible to compare and switch providers without adapting the code each time. Similar to OpenRouter, with a stronger focus on the whole range of AI services (see LLM gateway).

LM Arena (formerly Chatbot Arena) is a platform where users compare and rate the answers of two anonymous models. The votes produce a public leaderboard.

It is a helpful orientation but does not replace a test with your own tasks: a top spot says little about how well a model fits your use case (see benchmark).

n8n is an open-source workflow automation platform — comparable to Zapier or Make, but self-hosted and significantly more powerful for complex AI workflows. With n8n, we connect LLMs, databases, APIs, email systems and hundreds of other integrations into intelligent, automated pipelines.

The decisive difference from SaaS alternatives: n8n runs on your own infrastructure — all data remains under full control. This is especially important for GDPR-compliant AI automation. The visual node editor makes complex multi-step workflows configurable without programming skills.

vona uses n8n in projects where processes can be connected quickly, for example: for AI-assisted document processing, automated email workflows, CRM integrations and custom AI agents.

Zapier connects thousands of web applications through a building-block editor: an event (trigger) in one app sets off actions in others, without you having to code. AI steps can now be added as well.

Getting started is very easy; operation runs in the provider's cloud. For large volumes or sensitive data, it is worth comparing with self-hosted solutions such as n8n.

Make (formerly Integromat) is a visual automation platform: you drag modules onto a canvas, connect them and let processes between applications run automatically.

Compared with simple if-this-then-that tools, it is well suited to branching processes with several steps. Operation here also runs in the cloud, and AI building blocks can be connected.

Activepieces is an open-source platform for workflow automation that you can host yourself. It offers a visual editor and building blocks for many services, including AI applications.

Anyone who wants to keep control of their data finds an alternative to pure cloud services here — as with n8n.

Flowise is an open-source toolkit for assembling AI applications by drag and drop: chatbots, RAG systems and agents from prebuilt blocks.

It is based on libraries such as LangChain and is good for trying out ideas quickly. For long-term operation you should plan for hosting, updates and access rights.

Langflow is a visual editor for AI workflows and agents in the LangChain ecosystem. You connect components such as models, prompts and data sources on one canvas.

It targets teams that want to build prototypes quickly, even without deep programming knowledge, and can be extended with your own code when needed.

Relevance AI is a platform where companies build AI agents and deploy them in teams — a kind of digital coworker for recurring tasks such as research, sales or support.

The platform is operated as a service. Before using it, you should clarify which data is processed, where it is stored and how approvals and control are handled.

Bardeen automates tasks in the browser: take data from web pages, write it into spreadsheets and CRM systems, shorten recurring click sequences — increasingly with AI support.

Handy for individuals and small teams, for example for research and contact lists. When taking data from web pages, mind the terms of use and data protection.

OpenClaw is an open-source project for a personal AI assistant or AI agent that runs on your own hardware or server. It connects language models (for example from OpenAI, Anthropic or locally run models) with messaging channels and tools, takes tasks via chat and carries them out on its own — for example answering messages, managing appointments or doing research.

Typical for such agents are a memory across conversations, extensible capabilities (plugins or “skills”) and access to files, accounts and programs. That is exactly where the risk lies: the more rights an agent gets, the more important are clear permissions, an isolated environment, logs and approvals for critical actions.

Our view: Such agent platforms are interesting for companies but need security, privacy and permission concepts before they run in production. We advise on this and build agent solutions for clients with minimal rights, approval steps and traceable logs — with the tool that fits the task.

LangChain is a widely used open-source framework for building applications with language models: it provides building blocks for prompts, data sources, tools and agents in Python and JavaScript (see agent framework).

It connects models from different providers with your own data and tools. Because of its breadth and rapid development, it pays to use only the parts you really need.

LangGraph is a framework in which agents are described as a graph of steps and decisions — with state, loops and branches.

This allows multi-step processes to be controlled in a more disciplined way, with checkpointing and approval steps (human-in-the-loop). More suited to developer teams than to a quick prototype.

LlamaIndex is a framework that connects language models with your own data: documents are ingested, prepared, indexed and made searchable for questions (RAG).

It comes with connectors for many data sources and vector databases and is often used for knowledge assistants and document search.

AutoGen is an open-source framework from Microsoft Research for systems in which several AI agents communicate with one another and solve tasks together.

Typical are splits into roles, such as planner, programmer and reviewer, who pass results to each other. Such multi-agent systems are powerful but harder to test and control. The project status changes; check the current documentation before using it.

CrewAI is a framework for assembling teams of AI agents with clear roles, goals and tasks that work out a result together.

The approach is easy to grasp — a "researcher" supplies material, an "editor" writes it up — but it needs good task descriptions and a check of the results.

GPTs are customized versions of ChatGPT: with instructions, uploaded knowledge and optional connections to external services, you build an assistant for a specific task — without programming.

They can be shared with a team or published. A quick route for simple, recurring tasks; with complex processes or special data protection requirements you hit limits.

Copilot Studio is Microsoft's platform for building your own copilots and agents that access company data and trigger processes in the Microsoft environment.

Its advantage lies in the integration with Microsoft 365, Teams and the company's permission management. Anyone who mostly works in that world gets to results quickly.

AutoGPT and BabyAGI were among the first widely noticed experiments with autonomous AI agents in 2023: you give a goal, and the system breaks it into tasks and works through them on its own.

Today they mainly serve as a reference for how agent loops are built. For productive use, controlled frameworks and clear approvals are the better choice.

Claude Code is Anthropic's coding agent for the command line. It reads your project, proposes changes, runs commands and tests and works through tasks step by step — with your approval.

vona uses Claude Code in everyday development, for example for refactoring, tests and documentation. What stays important: results are reviewed before they are adopted into a project.

Replit is a browser-based development environment where you program, run and publish applications — without any local setup.

With the built-in AI agent, working apps can be generated from a description. That suits prototypes and ideas well; for production software, the generated code should be reviewed and reworked.

Kiro is a development environment with AI agents built by Amazon Web Services. Its approach: requirements and design are first captured as a specification, then the agent implements them (spec-driven development).

This is meant to make AI-assisted programming more traceable than pure "vibe coding" by shouting orders. As with all coding agents: the generated code has to be reviewed.

Devin is an AI software agent from the company Cognition that works on programming tasks largely on its own — with its own development environment, terminal and browser.

Such agents take on clearly scoped tasks such as bug fixes or smaller features. People review the results (code review) before anything goes into production.

E2B provides isolated environments (sandboxes) in which AI agents can run code without endangering your own system.

This matters as soon as an agent writes and launches programs itself: errors or unwanted commands stay inside the sandbox. It is connected via API.