Eulogik

We engineer hyper-efficient, domain-specialized, and edge-deployable AI models.
No bloated parameter sizes. No massive cloud dependencies. Just high-performance models designed to run locally, securely, and cost-effectively.

🔬 Active Research & Model Portfolios

We focus on three primary domains of edge-optimized machine intelligence:

1. 📄 Visual Document Understanding (VDU)

TinyDoc-VLM-256M: A compact 290M-parameter vision-language model trained on SigLIP and SmolLM2. Optimized for document OCR, layout parsing, and information extraction.
TinyDoc-VLM-LoRA: Parameter-efficient adapters for custom corporate legal and administrative document compliance.

2. 📈 Time Series & Quantitative Forecasting

NanoForecast: A lightweight time-series foundation model family ranging from 200K to 6.5M parameters. Supports streaming RNN inference and zero-shot forecasting, optimized for running directly on edge hardware like Raspberry Pi.

3. 🐡 Intelligent LLM Routing

fugusashi: An open-source, federated, human-interpretable model router. Learns user preferences and routes requests to the most optimal model, slashing inference costs by up to 70%.

4. 🇮🇳 Localized & Multilingual LLMs

Bharat-Tiny-LLM: An ultra-compact 240M-parameter language model specialized for Hinglish and Indian cultural context.

⚡ Developer Demos (Spaces)

Test our models directly in your browser:

🚀 Bharat-Tiny-LLM Chatbot — Hinglish AI companion.
🐡 Fugusashi Router — Intelligent prompt router.
📄 TinyDoc-VLM Demo — Local document analysis.
📊 NanoForecast CSV Tool — Zero-shot time series analysis.

🏢 Licensing & Enterprise Collaboration

All our core public models are open-sourced under Apache 2.0 or MIT licenses.

For enterprise-grade integration, private custom model training, or commercial licensing of specialized components (such as our high-frequency quant trading predictors and legal document ingestion pipelines), get in touch: