Modernize legacy codebases faster with AI-guided migrations tailored to your needs.

In Java, you can upgrade your codebase using OpenRewrite for rule-based changes and LLM tooling for error correction, running automatically in incremental steps. If you prefer more autonomy, a fully LLM-driven flow performs small, developer‑supervised migration steps.

For complex legacy languages like PL/I, our custom parser builds an abstract syntax tree that supplies context to LLM-based transformations, facilitated by a custom IntelliJ plugin.

You can find further information also in our LinkedIn articles about Java and PL/I Migration.

Step into the AI Photobooth and become the stars of your own image!

Gather up to three friends and take individual photos. Then, choose from a variety of themes or mix and match your favorites. Our AI will seamlessly blend your portraits into a brand-new scene, creating a unique and personalized image. Once your masterpiece is ready, print it out and take home a one-of-a-kind souvenir.

Our AI Photobooth showcases what's possible with state-of-the-art image generation models.

Tune in to the TNG AI Radio Station and listen to a broadcast built around what you want to hear!

Enjoy a playlist curated to your taste and add what you care about with individually configurable content such as news or local weather. Our AI pulls updates from the internet, summarizes them, and delivers them through lifelike, AI-generated host voices - anytime, whether in the car, at home, or at work.

Hold a presentation with zero preparation as AI Slides builds and delivers your slides!

Simply provide a title and AI Slides will create an entire slide deck, all set up with images, headlines, and speaker notes. When it’s time to present, all you have to do is take a quick snapshot of your face, record a few seconds of your voice, and your digital clone will deliver the talk.

AI Slides combines the spear head of GenAI models. Notably, the virtual avatars are generated by D-ID and voiced by an ElevenLabs Voice Clone.

Ever wondered what Vincent Van Gogh might have envisioned when he painted "Starry Night"? Experience it firsthand with ArtVision.

ArtVision turns the real world into an artistic painting with the help of AI and Extended Reality. Just slip on the VR goggles and dive into our vast selection of famous paintings – and some less famous ones.

The showcase is based on “neural style transfer”, a technique that enables the underlying neural network to apply an image’s style to the headset’s live camera feed.

Boost your productivity with our innovative Atlassian apps!

TNG is an Atlassian Silver Marketplace Partner dedicated to developing practical apps for Jira and Confluence. We empower teams with state-of-the-art technology to improve efficiency and productivity. At Big Techday 25, we’re showcasing four of our sixteen powerful apps:

Rock your Sprint Review for Jira

Transform your agile sprint reviews with AI-driven automation.

This app generates polished, insightful presentations directly from your Jira sprint data, intelligently summarizing and organizing issues. Enhance slides with DALL·E-generated visuals or your own screenshots for a personalized touch. Fully integrated into Jira, it saves time and improves stakeholder communication making every sprint review more effective.

Multivote & Enterprise Survey for Confluence

Simplify event planning and meeting coordination with dynamic, multi-user surveys and no coding required!

You can create conditional, interactive surveys directly in Confluence, collect real-time feedback, and allow participants to update their responses until final submission. With its seamless integration and instant updates, our app ensures accurate, efficient decision-making.

Next Generation Translator for Confluence

Break language barriers and collaborate globally!

With our Next Generation Translator, you can instantly translate Confluence pages to your preferred language. The app remembers your settings and can be integrated with leading translation services or your own backend to ensure full data privacy and compliance. Whether for teams or enterprises, it enables seamless multilingual collaboration right within Confluence.

Dark Mode & Rotation for Confluence Images

Optimize visual content for every user experience!

This app automatically adapts images to dark mode or lets you specify separate images for light and dark themes, displayed based on the user’s browser settings. Easily rotate images in-place without having to re-upload them - making image management in Confluence smoother than ever.

CaptAIn Bookworm is your bedtime co‑author that never gets bored, runs out of breath or stuck on page two.

Children’s books build empathy, creativity, and understanding - but parents know the “again, again!” loop all too well. CaptAIn Bookworm is an AI‑powered story generator that spins personalized, heart‑warming tales starring your kid’s imaginary friends. It delivers lovingly crafted stories with matching illustrations and a natural read‑aloud - voices, pauses, and all the cozy vibes.

Under the hood, the showcase uses conditioning to write coherent, age‑appropriate narratives that adapt to inputs like your child’s name and favorite characters, and to generate illustrations with characters that remain consistent across pages.

Turn chess games into lessons!

Our Chess AI generates narrated descriptions with explanations, plans, and puzzles from real-life games.

Engines like Stockfish or AlphaZero now far surpass humans, yet the educational aspect for average players is largely unexplored. We combine evaluations of state‑of‑the‑art neural‑network engines with traditional methods to create text via GPT‑4, then a chess instructor narrates with ElevenLabs. Our YouTube Shorts from online tournament games have 175k views. New Shorts are published daily on our YouTube channel @TNGAIChess.

What happens when you combine three popular open-source LLMs into one fast, sharp model? You get our DeepSeek-TNG R1T2 Chimera!

Released in July 2025, TNG’s R1T2 Chimera is a tri-mind Assembly-of-Experts LLM merging DeepSeek’s R1-0528, R1, and V3-0324. Built in under an hour of CPU time via our linear-time Assembly of Experts method, it balances intelligence and efficiency, running ~20% faster than R1 and more than twice as fast as R1-0528. It also outperforms R1 in benchmarks like GPQA-Diamond and AIME-24/25, improves <think>-token consistency, and is perceived as having a polite persona. The Model Weights are available on Hugging Face under the MIT license.

The EX-1Evo represents the latest stage in the development of solid-fuel rockets within WARR. Compared to its predecessor, the EX-1E, its structural weight has been reduced by 38%. Thanks to the additional use of a more powerful engine in the upper stage, the EX-1Evo reached a speed of Mach 2.5 at an altitude of 6,100 m at the International Rocket Engineering Competition. At Big Techday 25, the complete booster stage and the remains of the second stage after an anomaly shortly after burnout are on display.

Meet G1PO, TNG’s humanoid robot with eyes, purpose-built joints and the work ethics of Wall-E!

Our G1PO offers a glimpse into a future where robots learn from, work alongside and grow with us. It walks on its own using cameras, depth perception, and LiDAR to figure out where it is and where to go next. We trained it with reinforcement learning for walking and imitation learning for more complex behaviors. And yes, it can even double as a go-kart chauffeur, offering free rides to children.

Get ready for a family-friendly AI showdown!

In each round, a random prompt from GPT4 constructs an image that is generated by Stable Diffusion XL. Your task? Come up with a prompt that produces an image closest to the original. The more creative and accurate you are, the better your chances of winning.

Points will be awarded based on the degree of similarity between your image and prompt and the original content, as determined by AI metrics. The player with the most points takes the win! Will you be the champion of this AI-powered game?

What's the best LLM for agentic coding? How about tool calling? And what's the best LLM for extracting text from German invoices? LLM Benchmarking gives you the knowledge to decide how to pick the right LLM for your use case!

We built a toolbox for testing LLMs with a selection of meaningful, state-of-the-art benchmarks. Our framework allows for testing proprietary and open-weights LLMs through API calls.

We offer proven, community-driven benchmarks that enable like-to-like comparison to third party results, such as SWE-bench and Humanity's Last Exam. We also built custom benchmarks for our own, real-life use cases, covering blind spots in public benchmarks. This allows us to see the full picture of LLM performance ranging from agentic coding, math, multi-language understanding, answering tricky questions in STEM with public benchmarks all the way to German OCR, tool calling and visual understanding.

MaRLOS (Machine-Regulated Logic Operating System) is our friendly robot for maintaining justice and order in the TNG Maker Space. Inspired by Valve's GLaDOS from the Portal games series, it monitors the room with the patience of a saint and the judgemental gaze of an all-seeing AI.

The brain of MaRLOS is an LLM agent system that uses a Vision Language Model to understand and analyze the world around it. This allows MaRLOS to detect rule breaches and tell the violator off either via slack message or through the intercom.

Turn messy documents into clean, structured text with our fine-tune of olmOCR!

Built on the Allen Institute for Artificial Intelligence’s olmOCR, we fine-tuned a faithful Optical Character Recognition (OCR) engine for document parsing. While olmOCR's primary purpose is generating training data for LLMs through text linearization into logical reading order, it intentionally omits headers and footers as these cannot be meaningfully predicted during next-token prediction training. Our model specifically keeps these elements enabling a broader range of applications, including invoice parsing and more.

Tired of seeing the same old face in the mirror? Our Realtime Deepfakes got you covered!

Realtime Deepfakes swap out the user's face with the help of deep neural networks. The appearance of the entire head is changed while facial expressions are preserved. The result is an almost perfect illusion that has been optimized to run live in real-time. Thanks to careful engineering, this works for any person that steps in front of the camera, no matter the age, gender or ethnicity.

The showcase combines state-of-the-art computer vision models for pose detection, landmark detection, face segmentation with generative models for neural inpainting and face generation.

Playing board games without friends has never been easier. Well, almost. While our robots aren't quite ready to roll the dice with you, they've already mastered the art of sorting with style.

This showcase features a robotic arm setup with four SO101 arms (two leaders, two followers), powered by a custom finetune of the ISAAC-GR00T model, a cutting-edge foundation model for robot control built on a diffusion transformer architecture.

The system works fully end-to-end: camera images, joint angles and a prompt go in, an entire action chunk comes out. But generating full motion sequences in real time isn't easy. Without refinement, these chunked actions often lead to stuttery, unnatural movement.

Enter Real-Time Action Chunking: an extension inspired by visual inpainting, but for robot actions. Instead of generating all actions from scratch each time, we keep the first few fixed to the previous prediction and let the model "inpaint" the rest. The result? Far smoother, more natural motion.

Want to put your reflexes to the test? Step up to the controller, teleoperate one of the arms in real time, and challenge the robot in a 1 vs. 1 dice-sorting duel.

Keep your AI interactions private and fast with skAInet!

skAInet is TNG’s GPU cluster for running state‑of‑the‑art AI services. We own the stack - models and hardware - so confidential workloads and data stay in‑house with zero reliance on external providers. Across 8+ fully stacked GPU nodes, skAInet powers chat, agentic coding, image generation, transcription, translation, and many other tasks. Under the hood, dynamic batching, intelligent routing, and load‑balancing deliver low latency and high throughput for many parallel users.

Wish you could see your rough sketches come alive? Watch as Sketch2Image turns them into art!

Drop in a doodle, a stick figure, or even a messy napkin sketch. Sketch2Image transforms your input into rich, full-color artwork in whatever style you choose, be it soft watercolors, comic ink, painterly oil, photorealistic, or something entirely different. Sketch2Image preserves the structure of your original composition while filling in details and adding textures and colours.

Want the power of AI without giving up your privacy? TAIA is the tool of choice!

TAIA (TNG AI Assistant) is TNG's privacy first assistant that lets our colleagues work with large language models safely every day. It defaults to self hosted models, offers optional access to select commercial providers, and never logs any data. The UI makes model choices transparent, shows exactly who receives your data, and warns you whenever a request would go to a third party. Built-in tools include Internet Grounded Search, System Prompts, Artifact Creation, and more.

TAIA is also available to other companies as a ready to run solution, and we are happy to help you evaluate the fit.

Let AI write your tests – from unit to end‑to‑end.

UnitTestGen analyzes your business logic and repository context to generate working unit test suites, using live feedback from compiling and running the tests. The quality of the end result is validated using mutation testing coverage.

E2ETestGen uses an autonomous AI agent powered by a vision language model to interact with your web UI, navigating flows, completing forms, and checking outcomes. Afterwards, an LLM refines the recorded actions into maintainable Playwright tests.

Meet The Artist, a 1.3‑trillion‑parameter Assembly‑of‑Experts language model that thinks creatively without losing the plot!

In use, it delights with curious, out‑of‑the‑box ideas while keeping a firm semantic link to your instructions - hence its name. Technically, sparse expert routing selects a few specialists per token, combining originality with stability and speed. Explore prompts that mix style and rules; watch it riff, iterate, and still honor constraints for reliable, imaginative results.

There's a lot more to AI than Large Language Models. Digitalize a document scan, transcribe a meeting recording, or generate an image for your presentation with just a few clicks in your browser with the TNG AI Services!

No GPU? No problem! The heavy lifting of running countless AI models is taken care of quickly and securely by skAInet, TNG's own AI Cluster.

TWAIN transforms how TNG colleagues interact with their digital workspace, making tedious tasks conversational, all while adhering to privacy-first principles.

TWAIN, the TNG Web AI Navigator, acts as an intelligent browsing assistant that seamlessly navigates internal company tools and the open web. Beyond typical chatbot capabilities, TWAIN autonomously completes complex workflows, from booking desks and managing timesheets to ordering office supplies on Amazon Business. It reads emails, analyzes Confluence pages, and even fills out forms without constant user guidance.

Need a hand? The UR5 by Universal Robots is an industrial robotic arm with a payload of 5 kg. With its powerful gripper and six degrees of freedom, it can grip and snip just about anything! In the past, programming such a robot arm relied heavily on hard-coded motion sequences or complex computer vision and motion planning pipelines. Today, Robot Foundation Models like NVIDIA's GR00T N1 reduce the barrier of entry for solving complex tasks: In goes the camera feed and the instruction as text, out come fine-grained robot control instructions. Using the power of LLMs, they show best-in-class generalization and reasoning capabilities. New robot embodiments and manipulation tasks can be learned through small fine-tuning.

In this showcase, we fine-tuned GR00T N1.5 to control the UR5 for a color-based pick-and-place task. The robot receives a text prompt (e.g., “place the yellow cube on the green platform”) and performs this action, moving the cube onto the specified platform. The model was fine-tuned with imitation learning – exclusively on synthetic data simulated in Isaac Lab. Remarkably, despite only learning from simulation, GR00T is capable of transferring these capabilities to the real world.

Crysis VR

When Crytek's (in)famous first-person shooter Crysis was first released in 2007, it was so hardware hungry that even the most high-end PCs of the time struggled to run it. For years, any new PC hardware ultimately had to face the question "But can it run Crysis?", turning the game into a bit of a meme. 15+ years later, hardware has evolved significantly, and any issues running the game are more likely coming from its SecuROM copy protection causing problems rather than the hardware not being powerful enough.

It is time to up the ante: Let's run Crysis in Virtual Reality! With a game modification and some VR glasses, it is now possible to experience the classic from a whole new dimension. See the game's beautiful landscape around you rather than just as a projection on a screen in front of you. You can freely look around and get a glimpse of what it might actually feel like to be stranded on a tropical island as a covert super soldier.

Half-Life 2 VR

Valve's Half-Life is one of the most well-known video game franchises in the world. Its second installment, originally released in 2004, sees players take on the role of crowbar-wearing physicist Gordon Freeman to combat an alien invasion. The game was well received by players and critics alike, praising its diverse environments and gameplay and its clever and innovative use of physics. To this day, it remains one of the highest rated games of all time.

Almost 20 years later, a fan-made modification for Half-Life 2 allows you to experience the entire game in Virtual Reality. Wearing a VR headset and wielding motion controllers, you can step into Gordon's shoes like never before, actually seeing the world from his perspective and physically swinging the crowbar around.