Projects

Overview: This project is a deep research workflow implementation on the Dify platform, focusing on cascaded multi-source search, outline generation, and long-form report creation. It uses a modular design and supports flexible model and data source substitution. Key features: - Intelligent theme parsing using Gemini 2.0 Flash model for multi-level topic decomposition. - Mixed search engine combining local knowledge base and multiple external sources (e.g., Wikipedia, Google, Bing API). - Dynamic rhythm control using a 2>1 model cascade architecture with conditional branches and dialogue rounds for optimizing processing rhythm. - Efficient report generation with deep-seek-r1-distill models and Markdown structured output. Tech stack: Python, Dify, Gemini 2.0 Flash, deep-seek-r1-distill, and other related technologies. Why it matters: This workflow is significant as it provides a comprehensive solution for generating structured research reports, leveraging multiple sources and advanced models to achieve efficient and effective research.

### Overview: The AI Real Estate Assistant is an open-source project that leverages advanced AI techniques such as RAG, vector search, and multi-provider language models to provide property search, analytics, and market insights. It features a conversational interface, live demo, and supports multiple languages, making it accessible and useful for real estate enthusiasts and professionals alike. ### Key Features: - **Natural Language Property Search**: Ask questions like "2-bedroom apartment in Kraków under 500k" and get matched listings. - **AI-Powered Chat Interface**: Conversational experience for finding your perfect property. - **Financial Tools**: Mortgage calculator, rent-vs-buy comparison, ROI analysis, and TCO calculator. - **Interactive Maps**: Clustered property markers with area analytics. - **9 Languages**: English, Polish, Russian, German, Spanish, Italian, Portuguese, Turkish, and Ukrainian. ### Tech Stack: - Python - FastAPI - Next.js - ChromaDB - Multi-Provider Language Models ### Why It Matters: The AI Real Estate Assistant offers a seamless and intuitive way to search for properties, analyze market trends, and make informed decisions. Its ability to provide accurate property value estimates and financial tools makes it a valuable resource for both buyers and sellers in the real estate industry.

### Overview: OrbitKV is a high-performance key-value cache designed for large language models (LLMs) and serverless computing. It extends the KV cache capabilities of vLLM and SGLang beyond GPU memory, allowing for the reuse of prefixes in DRAM and optional SSD, and restoring them after GPU eviction or engine restarts. This feature is particularly useful for serving repeated documents, shared system prompts, and long conversation prefixes. ### Key features: - **Reuse beyond HBM:** Store prefixes in DRAM and optional SSD, and restore them after GPU eviction or engine restart. - **Share cache capacity:** Connect multiple local engines to one Manager or fetch compatible peer replicas through Mooncake TENT TCP/RDMA. - **Rust ownership:** Native cache and transfer owners manage byte budgets, reservations, and completion; Python connects engine callbacks and layouts. - **Local discovery:** Bounded inventory streams maintain covered indexes; ordinary queries avoid per-key etcd calls and validate source grants. ### Tech stack: Rust ### Why it matters: OrbitKV is essential for optimizing the performance and efficiency of large language models and serverless systems, particularly in scenarios where GPU memory is limited. By leveraging DRAM and SSD for cache storage and transferring data, it significantly reduces the memory footprint and improves the overall runtime of the models. ### Repository: Repository: OrbitInfra/OrbitKV GitHub stars: 319 ### Primary language: Rust ### GitHub description: High Performance KV Cache for LLM Inference ### Get started: Follow the installation and quickstart guide

### Overview: This repository compiles a comprehensive summary of existing multimodal question answering datasets and systems in the medical domain, focusing on both medical question answering (QA) and visual question answering (VQA) tasks. The datasets span various sources and challenges, providing valuable resources for researchers and practitioners interested in advancing medical NLP. ### Key Features: - **Medical QA Datasets:** - Corpus for Evidence Based Medicine Summarization (Mollá, 2010) - CLEF QA4MRE Alzheimer’s task (Peñas et al, 2012) - BioASK datasets (2012-2020) - TREC LiveQA-Med (Ben Abacha et al, 2017) - MEDIQA-2019 datasets on NLI, RQE, and QA (Ben Abacha et al., 2019) - MEDIQA-AnS dataset of question-driven summaries of answers (Savery et al., 2020) - MedQuaD Collection of 47k QA pairs (Ben Abacha and Demner-Fushman, 2019) - Medication QA Collection (Ben Abacha et al., 2019) - Consumer Health Question Summarization (Ben Abacha and Demner-Fushman, 2019) - emrQA: QA on Electronic Medical Records (Pampari et al., 2018) - EPIC-Q

### Overview: TableQA is an AI tool designed to query natural language on tabular data, utilizing QA models from Hugging Face. It supports querying from multiple CSVs, handling fuzzy string implementation, and supports databases such as SQLite, PostgreSQL, MySQL, and Amazon RDS. Open-domain, no training required, and customizable with manual schema and auto-generated schemas, making it an efficient and user-friendly solution for data querying. ### Key Features: - Supports detection from multiple CSVs (including Amazon S3) - FuzzyString implementation for incomplete column values - Supports databases: SQLite, PostgreSQL, MySQL, Amazon RDS - Open-domain, no training required - Add manual schema for customized experience - Auto-generate schemas in case of a missing schema - Data visualizations ### Tech Stack: Python ### Why it Matters: TableQA simplifies the process of querying tabular data with natural language, making it accessible and efficient for users without the need for extensive training. Its ability to handle multiple databases and support for manual schema makes it versatile and user-friendly, ideal for various applications requiring data querying.

### Overview Project Insight is designed to provide NLP as a service, utilizing Python code bases for both the front-end GUI (Streamlit) and the backend server (FastAPI). It supports a variety of downstream NLP tasks such as News Classification, Entity Recognition, Sentiment Analysis, and Summarization. The solution is expandable and can be easily updated and maintained, thanks to its microservices architecture. ### Key Features - **Python Code Base**: Built using FastAPI and Streamlit, making the complete code base in Python. - **Expandable**: The backend is designed to be easily expanded with more Transformer-based models, which will be automatically available in the front-end app. - **Micro-Services**: The backend is built with a microservices architecture, using Dockerfiles for each service and leveraging Nginx as a reverse proxy for independent services. ### Tech Stack - **Python**: The backend and frontend are built using Python. - **FastAPI**: The backend is built using FastAPI for microservices architecture. - **Streamlit**: The front-end is built using Streamlit for a user-friendly interface. ### Why It Matters Project Insight simplifies the process of NLP as a service by providing a Python-based code base for both the frontend and backend. This makes it easy to update, maintain, and run individual NLP services independently. The expandable nature of the backend allows for easy addition of new models, ensuring that the solution remains versatile and adaptable to changing needs.

Overview: This repository is a curated collection of AI tools designed to help developers streamline their workflows, automate tasks, and enhance their projects. It features a variety of tools from different categories such as content creation, monetization, development, automation, markets, and voice communication. Key features: - ViralWave Studio for creating and scheduling social content - HeyGen for generating polished AI videos - Authflow for monetizing a custom GPT - EverBee for researching and selling products online - Lovable for building an app or website with AI - Make.com for automating a workflow - CoreClaw for extracting structured public web data - Apify for monitoring websites and feeds - KamdenAI for following a structured morning market workflow - ElevenLabs for generating realistic AI voice - Chatbase for deploying an AI customer-support agent - CallHippo for automating business communication Tech stack: Unknown Why it matters: This repository provides a comprehensive collection of AI tools that can significantly enhance the development and business processes of developers, making their projects more efficient and effective.

Overview: This tutorial is designed for developers interested in deep learning and provides a comprehensive guide to the International Summer School on Deep Learning, 2019. Key features: - Pre-trained word embeddings - Pre-trained contextual embeddings - Fine-tuning a pre-trained model Tech stack: Jupyter Notebook Why it matters: This tutorial offers practical insights into deep learning concepts and hands-on experience with various NLP tasks, making it valuable for developers looking to enhance their skills in the field. Repository: dpressel/dliss-tutorial GitHub stars: 320 GitHub description: Tutorial for [International Summer School on Deep Learning, 2019](http://dl-lab.eu/) in Gdansk, Poland README excerpt: # dliss-tutorial Tutorial for [International Summer School on Deep Learning, 2019](http://dl-lab.eu/) in Gdansk, Poland ## Sections ### Overview Talk https://docs.google.com/presentation/d/1DJI1yX4U5IgApGwavt0AmOCLWwso7ou1Un93sMuAWmA/ ### Tutorial There are currently 3 hands-on sections to this tutorial. - The [first section](1_pretrained_vectors.ipynb) covers pre-trained word embeddings [(colab)](https://colab.research.google.com/github/dpressel/dlss-tutorial/blob/master/1_pretrained_vectors.ipynb) - The [second section](2_context_vectors.ipynb)

Overview: LightNet is a Python interface to DarkNet, a neural network library known for its state-of-the-art object detection models, particularly YOLO and YOLOv2. It simplifies the use of DarkNet for developers. Key features: - Easy-to-use via Python - State-of-the-art object detection - Lightweight and self-contained - Easy to install - Cross-platform - 10x faster on CPU Tech stack: CPython 2.7, 3.5, 3.6. Only 64 bit. Why it matters: LightNet simplifies the use of DarkNet for developers, making it easier to integrate object detection and image segmentation features into projects.

**Overview:** Fast and Portable Character String Processing in R (with the Unicode ICU) **Key features:** - String concatenation, padding, and wrapping - Substring extraction - Pattern searching (Java-like regular expressions) - Collation and sorting - Random string generation - Case mapping and folding - String transliteration - Unicode normalisation - Date-time formatting and parsing **Tech stack:** C++ **Why it matters:** A comprehensive tutorial and reference manual is available at <https://stringi.gagolewski.com/>. The package is designed for fast, consistent, and portable string processing in R, leveraging the International Components for Unicode library.

### Overview: KOMORAN 3.0 is a Korean Morphological Analyzer implemented in Java. It is a lightweight, independent, and easy-to-use tool for analyzing Korean morphology. ### Key Features: - **Pure Java Implementation**: KOMORAN is written entirely in Java, making it compatible with any Java environment. - **Library Independence**: It uses only the libraries developed by the developer, reducing external library dependencies. - **Memory Efficiency**: The tool can process morphological analysis with a 50MB memory footprint, making it suitable for small to medium-sized applications. - **Readable and Editable**: The format of the input text is easily readable and editable, making it a user-friendly tool. - **Flexible Output**: It provides analysis results in a format that includes spaces, allowing for easy parsing and further processing. ### Tech Stack: Java ### Why it Matters: KOMORAN is a powerful tool for analyzing Korean morphology, making it an essential resource for researchers, language enthusiasts, and developers working with Korean language data. Its lightweight nature and easy-to-use interface make it an ideal choice for various applications, including natural language processing, text analysis, and more. ### Repository: The repository is maintained by shineware and is available on GitHub at [https://github.com/shineware/KOMORAN](https://github.com/shineware/KOMORAN). ### Installation: To get started with KOMORAN, follow the installation guide provided in the README. ### Usage Examples: - **Example 1**: Analyze the sentence "

Overview: LLM-TPU is an open-source project designed to quickly deploy popular generative AI models like LLMs and VLMs on SOPHGO's TPU hardware, including BM1684X, BM1688, and CV186X chips. Key features: - One-click deployment of LLMs and VLMs on SOPHGO's TPU - Supports multiple models including Qwen3-Embedding, Qwen3-TTS, Mage-VL, Step3_VL, Falcon-Perception, LocateAnything, MiniCPM-V-4.6, Gemma4, InternVL3, Qwen2.5-VL, Qwen3, QwQ-32B, and DeepSeek-R1-Distill-Qwen-32B - Python and C++ demos available for various models - Supports cross-lingual reranking, text-to-speech, image understanding, and more Tech stack: C++ Why it matters: This project simplifies the deployment of cutting-edge generative AI models, making it easier for developers to integrate these models into their projects quickly and efficiently.

### Overview: Glyd is a lossless AI compression library that reduces GPU memory usage by 33% while maintaining bit-for-bit accuracy. This makes it ideal for models that need to run on smaller GPUs or handle more requests on the same GPU. ### Key Features: - **Lossless Compression**: Glyd stores weights and key-value pairs in fewer bits, ensuring that the model's outputs remain identical to the original. - **GPU Memory Reduction**: It achieves this by rebuilding the model's exact values as the model runs, using fewer bits than the original model. - **Versatile Usage**: Glyd can be used with various models, including Qwen/Qwen3.5-9B and Qwen/Qwen3.8-27B, and supports both vLLM and Python-based quantization. ### Tech Stack: Rust ### Why It Matters: Glyd is crucial for models that require minimal GPU memory, such as Qwen/Qwen3.5-9B and Qwen/Qwen3.8-27B, enabling them to run on smaller GPUs or handle more concurrent requests on the same GPU. This makes Glyd an essential tool for developers working on AI models that need to be efficient and scalable. ### Repository: Repository: surya-koritala/Glyd ### GitHub Stars: GitHub stars: 320 ### Primary Language: Primary language: Rust ### CI Status: CI: [![CI](https://github.com/surya-koritala/Glyd/actions/workflows/ci.yml

Overview: Olla is a high-performance, lightweight proxy and load balancer designed for efficient management of Large Language Model (LLM) infrastructure. It routes requests intelligently across multiple inference nodes, supports a wide variety of native endpoints, and provides unified model discovery and cataloging. Key features: - Intelligent routing - Automatic failover - Model discovery and unified model catalogues Tech stack: Go Why it matters: Olla simplifies the management of LLM infrastructure by providing a robust, low-overhead solution that routes requests efficiently and seamlessly integrates with various LLM models and providers.

Overview: TalkingHeads is a Python library designed to facilitate communication with various chat agents, including ChatGPT, Claude, Copilot, Gemini, LeChat, HuggingChat, and Pi. It automates interactions with these chat platforms, making it easier to integrate them into workflows. Key features: - Seamless interaction with multiple chat agents - Login functionality for enhanced security - Support for custom interactions and plugins - Saving conversations in various formats for easy access - Ability to switch between different models Tech stack: Python Why it matters: TalkingHeads simplifies the process of integrating chat agents into workflows, making it easier to automate and streamline communication with these tools.

Overview: This project aims to develop a working Speech to Text module using Mozilla DeepSpeech, a state-of-the-art open-source automatic speech recognition (ASR) toolkit. The project is based on a paper published at KONVENS 2019. Key features: - DeepSpeech implementation - Support for German language - Uses Google's TensorFlow for easier implementation - Includes a speech corpus and language model Tech stack: Python Why it matters: This project provides a practical application of DeepSpeech for German ASR, which can be used in various audio processing pipelines. Repository: AASHISHAG/deepspeech-german GitHub stars: 319

Overview: EfficientWord-Net is a hotword detection engine based on few-shot learning, designed for developers to add custom hotwords to their applications without additional charges. It uses Python and Google's TFLite for real-time inference and is inspired by FaceNet's Siamese Network Architecture. Key features: - Few-shot learning for hotword detection - Purely written in Python - Supports Google's TFLite for faster inference - Inspired by FaceNet's Siamese Network Architecture - Performs best with 3-4 hotword samples collected directly from users Tech stack: Python, Google's TFLite Why it matters: EfficientWord-Net simplifies the process of adding custom hotwords to applications, making it an essential tool for developers looking to integrate voice recognition functionalities without the need for additional charges.

### Overview: This package provides evaluation and data-science capabilities for Presidio and PII detection models. It includes a fake data generator for generating synthetic sentences based on templates and fake PII, and supports the evaluation of Presidio's PII detection models and custom PII recognizers. ### Key features: - Fake data generator for PII recognizers and NER models - Data representation layer for data generation, modeling, and analysis - Model/Recognizer evaluation for Presidio Analyzer and custom Presidio recognizers - Helper functions for results analysis ### Tech stack: Jupyter Notebook ### Why it matters: This package is essential for developers and researchers interested in developing or evaluating PII detection models, existing Presidio instances, or custom PII recognizers. It also aids in generating new data for Named Entity Recognition models, ensuring a comprehensive approach to data science and PII detection.

Overview: This repository contains example code for combining zero- and few-shot learning with a small annotation effort to obtain a high-quality dataset with maximum efficiency. Specifically, it uses large language models available from OpenAI to provide initial predictions, then trains a smaller, supervised model on a curated dataset. Key features: - Zero-shot learning - Few-shot learning - Curated dataset for high-quality annotations Tech stack: Python Why it matters: This approach allows developers to quickly gather a gold-standard dataset through zero-shot or few-shot learning, making it an efficient method for obtaining high-quality annotations.

Overview: PyATE is a Python implementation of term extraction algorithms, including C-Value, Basic, Combo Basic, Weirdness, and Term Extractor, using spaCy POS tagging. It aims to automate the extraction of terms from text data. Key features: - C-Value - Basic - Combo Basic - Weirdness - Term Extractor Tech stack: spaCy Why it matters: PyATE is designed to automate the term extraction process, making it easier and faster to analyze large amounts of text data. It is particularly useful for researchers and data analysts who need to extract relevant terms from unstructured text, such as scientific papers or news articles. GitHub stars: 319 Repository: kevinlu1248/pyate