# Tabularis AI

> Tabularis AI builds efficient, privacy-preserving AI solutions. Based in Tübingen, Germany. We specialize in high-quality synthetic data generation, small and efficient AI models optimized for on-premise and edge deployment, and privacy-first enterprise tools. Our models run on consumer hardware — no cloud dependency required.

## What We Do

- **Synthetic Data Generation**: Enterprise-grade synthetic data that preserves statistical properties and relationships of real datasets. Used for privacy-preserving analytics, data augmentation, model training, and compliance-safe data sharing. Dedicated page: https://tabularis.ai/synthetic-data
- **Custom AI Models**: Specialized models trained from scratch or fine-tuned for exact enterprise workflows. Classification, extraction, generation, and domain NLP, tuned for measurable task accuracy and predictable latency. Dedicated page: https://tabularis.ai/custom-models
- **Knowledge Distillation**: Compress frontier teacher models into fast, cheaper, deployable specialist models. Lower inference cost and latency while preserving task quality. Dedicated page: https://tabularis.ai/knowledge-distillation
- **Deploy Anywhere**: Run AI where your data lives. On-premise, private VPC, air-gapped, or edge deployment with full data sovereignty and production-grade inference. Dedicated page: https://tabularis.ai/deploy-anywhere
- **Fast Data Labeling**: Automated data labeling for text, images, tables, time-series, audio, and RLHF. Combines foundation-model pre-labeling, active learning, and expert review to deliver training-ready labels quickly. Dedicated page: https://tabularis.ai/fast-data-labeling
- **Sentiment Analysis API**: Multilingual sentiment and emotion classification API for support tickets, reviews, surveys, social text, and customer feedback. Supports 23 languages, JSON labels and scores, batching, and a free developer tier. Dedicated page: https://tabularis.ai/sentiment-analysis

## Products

- [GReaT (be-great)](https://tabularis.ai/blog/introducing-great): Open-source Python library for generating realistic synthetic tabular data using language models. 140,000+ downloads, published at ICLR 2023. GitHub: https://github.com/tabularis-ai/be_great
- [Faust-1](https://tabularis.ai/blog/faust-1): 1.6B parameter German-first language model trained from scratch. Runs on consumer hardware (CPU, Apple Silicon, entry-level GPUs). Available on Hugging Face: https://huggingface.co/tabularisai/Faust-1
- [EU PII Safeguard](https://tabularis.ai/blog/eu-pii-safeguard): On-premise PII detection and redaction for 42 types of personal data across all 24 EU languages. 97% accuracy. GDPR, HIPAA & SOC 2 compliant.
- [GuideGen](https://github.com/tabularis-ai/guidegen): Constrained decoding library for structured JSON output from language models using Pydantic schemas.

## Key Pages

- Homepage: https://tabularis.ai
- Vision: https://tabularis.ai/vision
- Research: https://tabularis.ai/research
- Blog: https://tabularis.ai/blog
- Team: https://tabularis.ai/team

## Services

- [Synthetic Data Generation](https://tabularis.ai/synthetic-data): Realistic text, tabular, time-series, and image data for privacy-safe AI training, testing, and evaluation. GDPR-first by design.
- [Custom AI Models](https://tabularis.ai/custom-models): Trained or fine-tuned for specific workflows. Classification, extraction, generation, and NLP tasks tuned for measurable business outcomes.
- [Knowledge Distillation](https://tabularis.ai/knowledge-distillation): Compress frontier teacher models into fast, cheaper specialist students without losing task quality.
- [Deploy Anywhere](https://tabularis.ai/deploy-anywhere): On-premise, VPC, air-gapped, and edge deployment with containers, quantization, monitoring, and handover.
- [Fast Data Labeling](https://tabularis.ai/fast-data-labeling): Automated multi-modal labeling pipeline for training data. Supports text, images, tables, time-series, audio, and RLHF with foundation-model pre-labeling, active learning, and human verification.
- [Sentiment Analysis API](https://tabularis.ai/sentiment-analysis): API for multilingual sentiment analysis and emotion classification. Returns structured JSON labels, scores, credits, and latency metadata for product analytics, support routing, review analysis, surveys, and social monitoring.

## Blog Posts

- [Faust-1: A German-First Language Model](https://tabularis.ai/blog/faust-1): Introducing Faust-1, a 1.6B parameter language model trained from scratch for German, optimized for local deployment on consumer hardware.
- [YapBench: Do Chatbot LLMs Talk Too Much?](https://tabularis.ai/blog/yapbench): A benchmark for measuring LLM over-generation on simple prompts. Evaluation of 76 models reveals an order-of-magnitude spread in verbosity.
- [EU PII Safeguard](https://tabularis.ai/blog/eu-pii-safeguard): On-premise PII detection and redaction across all 24 EU languages with 97% accuracy. GDPR, HIPAA & SOC 2 compliant. No data leaves your infrastructure.
- [GReaT: Generating Realistic Tabular Data](https://tabularis.ai/blog/introducing-great): Open-source framework using transformer language models to generate high-quality synthetic tabular data. Published at ICLR 2023, 140,000+ downloads.

## Research Papers

- Do Chatbot LLMs Talk Too Much? The YapBench Benchmark (arXiv 2026): https://arxiv.org/abs/2601.00624
- Synthetic Data for Low-Resource Swahili Sentiment Analysis (AfricaNLP / EACL 2026): https://openreview.net/pdf?id=VQ0VRo4DGM
- Unlocking the Full Potential of Data Science Requires Tabular Foundation Models, Agents, and Humans (Dagstuhl 2025): https://openreview.net/pdf?id=aXMPvmBAm5
- Open Artificial Knowledge (ICML Workshop 2024): https://arxiv.org/abs/2407.14371
- Language Models are Realistic Tabular Data Generators (ICLR 2023): https://openreview.net/forum?id=cEygmQNOeI

## Contact

- Website: https://tabularis.ai
- Email: info@tabularis.ai
- Discord: https://discord.gg/7WqEKw652R
- GitHub: https://github.com/tabularis-ai
- Location: Tübingen, Baden-Württemberg, Germany