Projects

Overview: This repository is an exploration of Generative AI technologies and their applications. Key features: - LLMs (Large Language Models) - Frameworks - Langchain / Llamaindex - Vector DBs (Databases for storing and retrieving information) - RAG systems (Retrieval-Augmented Generation) - Evaluations, Monitoring, Observability - AI Agents - Graph RAG - Hybrid RAG - Performance Optimization - Production - Use Cases Tech stack: Jupyter Notebook Why it matters: This repository serves as a comprehensive resource for developers interested in exploring and experimenting with Generative AI technologies, providing insights into various frameworks, models, and best practices in the field. Repository: jayita13/GenerativeAI Primary language: Jupyter Notebook GitHub stars: 58

Overview: This project is a Pytorch implementation of the "Adaptive Co-attention Network for Named Entity Recognition in Tweets" paper presented at AAAI 2018. The code includes pretrained word vectors, extracted VGG features, and a re-implementation of the original paper's equations. Key features: - Pytorch implementation - Pretrained word vectors - Extracted VGG features - Re-implementation of the original paper's equations Tech stack: Python Why it matters: This project provides a practical and efficient way to perform named entity recognition in tweets using Pytorch, with improved performance compared to the original paper's implementation.

Overview: This repository is a frequently updated list of papers, models, and resources related to dLLM (Diffusion Large Language Models). It serves as an overview of the field, providing insights into the latest developments and resources. Key features: - Papers and models related to dLLM - Continuous Latent Diffusion Language Model - DiffusionGemma - dLLM - Continuous Diffusion Scales Competitively with Discrete Diffusion for Language - Causal Diffusion Language Models - Soft-Masked Diffusion Language Models - UltraLLaDA - Syntax-Guided Diffusion Language Models with User-Integrated Personalization - CoDA: Coding LM via Diffusion Adaptation - LLaDA-MoE: A Sparse MoE Diffusion Language Model - Dream 7B: Diffusion Large Language Models - Mercury: Ultra-Fast Language Models Based on Diffusion - LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models Tech stack: Python Why it matters: This repository provides a comprehensive and up-to-date overview of the latest developments in dLLM research, making it an essential resource for developers and researchers in the field.

Overview: Uni is a modern, embedded database that combines property graph, vector search, and columnar storage into a single, cohesive engine. Designed for applications requiring local, fast, and multimodal data access with object storage durability. Key features: - Embedded & Serverless: Runs as a library within your application — no server process. - Property Graph: OpenCypher queries with MATCH, CREATE, WHERE, ORDER BY, LIMIT, and aggregations. - Serializable Transactions: Snapshot isolation + optimistic concurrency control, on by default — conflicting concurrent commits abort with a retriable error instead of silently losing writes. - Locy Reasoning: Datalog-style recursive rules over the graph, with probabilistic semantics and neural predicates. - Vector Search: K-NN similarity search (L2, cosine) with pre-filter and threshold support, plus hybrid vector + BM25 fusion. - Columnar Storage: Lance-backed persistence on local disk or object storage (S3/GCS). - Graph Algorithms: PageRank, Louvain, shortest path, and more via the built-in algorithm library. - Forks: Named, durable, writable graph branches with nesting, TTL, structural diff, and content-UID-keyed write-audit-publish promotion to primary. See `website/docs/features/forks.md`. - Rust & Python: Native Rust crate and Python bindings (PyO3). Tech stack: Rust. Why it matters: Uni is designed for applications requiring local, fast, multimodal data access with object storage durability, making it suitable for a wide range of use cases, including applications that need to

### Overview: GuidedQuant enhances LLM quantization by integrating gradient information from the end loss into the quantization objective, boosting the performance of SOTA weight-only scalar, weight-only vector, and weight-and-activation quantization methods. ### Key Features: - **LNQ (Non-Uniform Scalar Quantization)**: Monotonically decreases the quantization objective value. - **End Loss Guidance**: Integrates gradient information from the end loss into the quantization objective. - **State-of-the-Art Performance**: Boosts the performance of SOTA weight-only scalar, weight-only vector, and weight-and-activation quantization methods. ### Tech Stack: Python ### Why it Matters: GuidedQuant is a significant advancement in LLM post-training quantization, offering improved performance and efficiency.

### Overview: This project is an end-to-end model for automatic speech recognition (ASR) using TensorFlow, trained on a small VoxForge dataset. It utilizes a CTC loss function and a 2-layer B-LSTM network to recognize speech in Russian. ### Key Features: - **Training on VoxForge Dataset**: The model is trained on a small dataset from VoxForge, specifically for the speaker "Aaron". - **Two Layers B-LSTM**: The model consists of two layers of Bidirectional LSTM (Bi-LSTM) with 128 neurons each. - **CTC Loss Function**: The model uses CTC (Neural Temporal Classification) loss function for training. - **MFCC Features**: The input features are extracted as Mel Frequency Cepstral Coefficients (MFCCs). - **Training and Testing**: The model is trained and tested using example files and a microphone for real-time speech recognition. ### Tech Stack: - **Python 2.7+** - **Jupyter Notebook** - **Tensorflow 1.0+** - **Keras** - **python_speech_features** - **numpy** - **scipy** ### Why It Matters: This project provides a practical solution for speech recognition in Russian, leveraging a small but effective dataset and advanced neural network techniques. It can be used for real-time speech recognition tasks, making it valuable for applications requiring accurate speech-to-text conversion.

Overview: Diff-eRank is a novel rank-based metric designed for evaluating large language models (LLMs) and multi-modal models. It quantifies how LLMs discard redundant information after training, providing insights into their performance and alignment quality. Key features: - Evaluates LLMs and multi-modal models using rank-based metrics. - Demonstrates applicability in both single-modal and multi-modal settings. - Consistently improves with model scaling, showing a relationship with traditional metrics like loss and accuracy. - Provides a method for assessing alignment quality in multi-modal models. Tech stack: Python, PyTorch, Transformers, and other relevant libraries. Why it matters: Diff-eRank offers a unique approach to evaluating LLMs and models, providing valuable insights into their performance and alignment, which can help researchers and practitioners better understand and improve model quality. GitHub stars: 58 Repository: waltonfuture/Diff-eRank Primary language: Python GitHub description: [NeurIPS 2024] A Novel Rank-Based Metric for Evaluating Large Language Models.

Overview: 让任何 AI 真正“看”视频：纯本地的视频阅读技能——faster-whisper GPU 转写 + ffmpeg 场景感知抽帧，双通道带时间戳，数据不出机。1 小时视频 3~5 分钟读完。 Key features: - 文字稿（带时间戳） - 关键帧（清晰度 512/768/1024，每张带时间戳） - 关键帧去重 - PDF 自动导出 - 本地与隐私保护 Tech stack: Python Why it matters: 为 AI 设计的本地视频阅读技能，无需依赖云端 API，数据完全在本地处理，提供高效、准确的视频阅读体验。 Repository: xiaohui5206/let-ai-read-video Primary language: Python GitHub stars: 56 GitHub description: 让任何 AI 真正"看"视频：100%本地视频阅读技能——faster-whisper GPU 转写 + ffmpeg 场景感知抽帧，双通道带时间戳（t=MM:SS），数据不出机。1 小时视频 3~5 分钟读完。

### Overview: This repository contains the official code for "One Noise to Rule Them All: Learning a Unified Model of Spatially-Varying Noise Patterns," authored by Arman Maesumi, Dylan Hu, Krishi Saripalli, Vladimir G. Kim, Matthew Fisher, Sören Pirk, and Daniel Ritchie. The project aims to develop a unified model for spatially-varying noise patterns, providing a comprehensive solution for generating noise textures. ### Key Features: - **Model Architecture:** The repository includes the complete model architecture, including the network blocks and hyperparameters. - **Data:** A procedural noise dataset is available for training and testing, consisting of 1.3 million textures with resolution 256x256 and noise parameters. - **Inference:** Scripts and examples for generating noise textures and inference processes are provided. - **Training:** Instructions for both single-GPU and multi-GPU training are included, along with detailed configuration files for Huggingface Accelerate. ### Tech Stack: Python ### Why It Matters: This project offers a significant contribution to the field of noise generation by providing a unified model that can generate spatially-varying noise patterns. The inclusion of a regularization term to enforce a spherical structure on the learned embedding space ensures that the model's outputs are consistent and can be used in various applications, such as texture synthesis and noise reduction in images.

### Overview: This repository contains a curated collection of creative prompts for OpenAI's GPT Image 2, featuring real previews generated via Atlas Cloud. The repository also includes a list of supported models and a guide on how to run any prompt via Atlas Cloud. ### Key Features: - **Creative Prompts**: A curated collection of prompts for GPT Image 2. - **Real Previews**: Generated via Atlas Cloud for a realistic preview. - **Model Support**: Supported models include Video, Image, 3D, LLM, Audio, and Explore more. - **Skill Integration**: Recommended to install [atlas-cloud-skills](https://github.com/AtlasCloudAI/atlas-cloud-skills) for Claude, Codex, and Gemini CLI. - **CLI Tools**: Available for CLI users, including [atlascloud-cli](https://github.com/AtlasCloudAI/cli). - **Free API Key**: Supports 300+ models with a single key, allowing access to all models. ### Tech Stack: TypeScript ### Why It Matters: This repository is a valuable resource for developers looking to enhance their AI creativity and explore new models. The curated prompts and real previews make it easy to experiment with GPT Image 2 and Atlas Cloud's capabilities.

### Overview: Chaty is a private desktop app for open models, including chat, coding agent, image studio, music studio, and document answers, all running on local models without the need for an account, cloud storage, or telemetry. ### Key Features: - **Code Mode**: The agent's actions are visible in the code editor, showing the changes made as they happen. - **Built for Small Models**: The model's training format is used for tool calls, ensuring that small models can handle the work. - **GGUF and MLX**: Both llama.cpp and MLX are native on every platform, and a model store searches Hugging Face for the best fit. ### Tech Stack: TypeScript ### Why It Matters: Chaty is designed to be a one-stop solution for a variety of tasks, all running on local models without the need for an account, cloud storage, or telemetry, making it ideal for users who prefer to keep their data and privacy intact.

### Overview: This repository contains the dataset and code for the paper "Language-driven Grasp Detection." The dataset is accessible via [this link](https://airvlab.github.io/grasp-anything/docs/download/), and the code is designed for training and evaluating various deep learning models for grasp detection. ### Key Features: - **Datasets:** The repository includes the dataset "Language-driven Grasp Detection" for training and evaluation. - **Training:** The default deep network used is GR-ConvNet. Users can train GR-ConvNet on different datasets using the provided command-line interface. - **Testing:** The repository also includes testing procedures for various baselines, allowing users to evaluate different models on the same dataset. ### Tech Stack: - **Python:** The primary programming language used for the implementation. ### Why It Matters: The repository provides a comprehensive solution for researchers and developers interested in improving grasp detection using language-driven methods. It offers a practical way to train and test various deep learning models on real-world datasets, making it a valuable resource for both academic and industrial applications. ### GitHub Stars: The repository has been starred by 57 developers, indicating its popularity and usefulness in the community. ### Repository: The repository is named "Fsoft-AIC/LGD" and is primarily written in Python. It has been starred by 57 developers, making it a valuable resource for researchers and developers in the field of computer vision and robotics.

Overview: This repository provides the recipe for the paper "A Crowdsourced Open-Source Kazakh Speech Corpus and Initial Speech Recognition Baseline" by offering a comprehensive Kazakh speech corpus, including high-quality transcribed data and additional sources, totaling around 1.2k hours of transcribed utterances. Key features: - High-quality transcribed Kazakh speech corpus - Over 600k utterances - Supplemental data from other sources Tech stack: Shell Why it matters: This repository serves as a foundational resource for researchers and developers interested in Kazakh speech recognition and language processing, providing essential data and tools for building and evaluating models. Repository: IS2AI/ISSAI_SAIDA_Kazakh_ASR GitHub stars: 57

Overview: ScrAIbe-WebUI is a user-friendly web service for automated transcriptions built on top of the innovative backend framework, ScrAIbe. It provides a no-code solution for deploying the service locally using Docker. Key features: - Gradio-Based Interface: Intuitive web interface for easy interaction without coding. - Real-Time and Asynchronous Transcription Modes: Perfect for live applications and asynchronous processing. - Comprehensive Media Support: Supports a wide range of file formats, including audio and video. - Advanced Model Integration: Multiple transcription models and speaker diarization. - Configurable and Extensible: Customizable settings and preferences via `config.yaml`. - Deployment Options: Docker compatibility and Docker Compose support. Tech stack: Python. Why it matters: ScrAIbe-WebUI simplifies the process of deploying transcription services, making it accessible to developers without extensive coding knowledge. The web interface, real-time transcription, and advanced model integration make it an efficient and powerful tool for transcribing audio and video files.

### Overview: The 2026 VRAM-Smart Fine-Tuning Suite for SDXL and ANIMA on 12GB GPUs is a precision-engineered environment that rethinks how gradient computation, optimizer states, and activation memory interact. It achieves a 12 GB VRAM baseline for SDXL-class models—equaling the quality of full-precision training on 80 GB enterprise cards—without the need for server clusters or cloud rentals. ### Key Features: - **Neuro-Compaction Layer**: Redistributes computational load intelligently, achieving a 12 GB VRAM baseline for SDXL-class models. - **Tensor Folding Engine**: Dynamically analyzes layer sparsity, activation frequency, and gradient flow to "fold" memory-intensive components into compressed, overlapping representations during the forward pass, then gracefully unfolds them for the backward pass only where absolutely necessary. - **Hybrid Momentum Distillation**: Combines the speed of low-order moments with the correctness of high-order corrections, reducing optimizer state memory by 60%. - **Sleep-State Gradient Manager**: Monitors the model's loss landscape per-layer, reducing precision and update cycle for layers in near-zero gradient states to prevent wasted computation. ### Tech Stack: - **HTML** ### Why It Matters: This is not a toy; it is a full-fidelity workshop for artists, researchers, and indie studios who demand studio-grade results without a datacenter budget. The suite is designed to meet the highest standards of generative AI while minimizing memory usage, making it ideal for artists, researchers

### Overview: PolyTask Orchestrator is a multi-task learning framework built on the modernBERT architecture, designed to streamline AI workflows by allowing researchers and engineers to fine-tune multiple tasks simultaneously, thereby improving accuracy across domains. ### Key Features: - **Task Synchronization Engine**: Automatically balances training updates across tasks to prevent one task from dominating the loss landscape. - **Loss Weight Autopilot**: Implements uncertainty-weighted losses to let the model decide which task needs more attention at each step. - **Task Grouping**: Cluster semantically related tasks to share low-level features exclusively among themselves. - **Dynamic Task Scheduling**: Automatically balances training updates across tasks to prevent one task from dominating the loss landscape. - **Loss Weight Autopilot**: Implements uncertainty-weighted losses (Kendall et al. method) to let the model decide which task needs more attention at each step. - **Task Grouping**: Cluster semantically related tasks (e.g., NER and POS tagging) to share low-level features exclusively among themselves. - **Frozen Layer Control**: Choose which transformer layers remain static (pretrained) and which become trainable (fine-tuned) on a per-task basis. - **ModernBERT Core**: Built on the latest efficient attention mechanism, supporting sequences up to 8,192 tokens without positional encodings collapsing. - **Modular Task Heads**: Each task gets a lightweight, task-specific head (linear, LSTM, CRF) that projects the shared representations. - **Zero-Shot Task Transfer**: Train on English tasks, evaluate on German or Mandarin—the shared

### Overview: This repository contains examples on how to use Tinkoff Voicekit, a service designed for developers to integrate voice recognition and synthesis functionalities into their applications. ### Key features: - **Voice Recognition**: Implementing speech recognition using Tinkoff Voicekit. - **Voice Synthesis**: Generating spoken text using Tinkoff Voicekit. - **Integration with C#**: Examples demonstrating how to integrate Tinkoff Voicekit into C# applications. ### Tech stack: C# ### Why it matters: These examples provide a practical guide for developers to integrate Tinkoff Voicekit into their applications, focusing on C# integration, which is essential for C# developers working on voice-based projects.

### Overview: RLLaVA is a user-friendly framework designed for multi-modal RL research, optimized for resource-constrained teams. It features an RL-central design, modular customization, and resource-efficient execution. ### Key Features: - **RL-Centric Design**: Decouples algorithm logic from distributed execution, enabling modular customization. - **Modular Architecture**: Enables easy development, extension, and customization of RL algorithms and multi-modal architectures. - **Resource-Efficient**: Optimized for resource-constrained setups, running most tasks on a single 24GB GPU. - **User-Friendly**: Minimalist code with familiar HuggingFace & PyTorch APIs for seamless setup and extensions. ### Tech Stack: - Python ### Why It Matters: - **Flexibility**: Unified plugin system for a variety of SFT-RL Fusion Algorithms. - **Accessibility**: Resource-efficient design makes advanced RL research more accessible to teams with limited resources. - **Ease of Use**: Minimalist code and familiar APIs make it easy to set up and extend experiments. ### Repository: - **GitHub Repository**: https://github.com/TinyLoopX/RLLaVA - **Stars**: 57 - **Description**: A framework for multi-modal RL research, optimized for resource-constrained teams. ### Installation: - **Clone Repository**: `git clone https://github.com/TinyLoopX/RLLaVA && cd RLLaVA` - **Set Up Environment**: `conda create -n rllava python==3.12 && conda activate rllava` - **Install

Overview: PolyglotMind is a comprehensive workshop for fine-tuning large language models, offering a modular training orchestration layer, multilingual support, and a transparent fine-tuning loop. It includes a responsive dashboard, multilingual tokenizers, a customer support channel, and a deterministic reproducibility ledger. Key features: - Modular training orchestration layer - Multilingual support with forty-plus tokenizer profiles - Multilingual datasets with right-to-left script handling, emoji-aware segmentation, and locale-sensitive normalization - 24/7 customer support - Deterministic reproducibility ledger - Adapter fusion studio for merging LoRA, prefix vectors, and prompt-tuning embeddings - Evaluation metrics including perplexity, BLEU, ROUGE, chrF, and a custom Semantic Drift Index - Offline-first operation with air-gapped environments and local sharding - Quantization ladder for bf16, int8, and 4-bit models - Privacy-preserving dataset vault with encrypted at rest and redaction filters Tech stack: Python, TensorFlow, PyTorch, and various NLP libraries Why it matters: PolyglotMind provides a robust, transparent, and extensible fine-tuning framework, making it ideal for practitioners looking to fine-tune models for multilingual and low-resource scenarios.

Overview: VisualClaw is an official implementation of a real-time, personalized agent for the physical world, designed to work seamlessly with various agent frameworks and downstream language models. Key features: - Gateway proxy for OpenAI/Anthropic-compatible APIs - Claude Code backend for tool-using agent scenarios - Codex backend for tool-using agent scenarios - Skill injection for retrieving relevant skills - Memory retrieval for storing and retrieving context - Self-evolution for skill bank evolution - Video cascade for handling video processing - Live vision for streaming camera frames - VisualClawArena packaging for benchmarking scenarios Tech stack: Python Why it matters: VisualClaw simplifies the integration of real-time, personalized agents with downstream language models, making it easier to develop and deploy personalized agents for various applications.