Projects

### Overview: Unfoldly is a local AI search tool designed to help you find files you can't remember by describing what you're looking for. It indexes your local files and allows you to search them by meaning, not just by filename or folder path. Unfoldly runs locally and doesn't require an account, keeping your files on your device. ### Key Features: - **File Search by Description:** Describe what you're looking for, and Unfoldly will find the right files. - **Local Memory Layer:** Your files are stored locally, and Unfoldly processes them to extract text, visual signals, metadata, transcripts, and searchable embeddings. - **Search and Ask:** Describe what you're looking for, inspect the source files Unfoldly finds, and ask follow-up questions in context. ### Tech Stack: Unfoldly is built using Python. ### Why It Matters: Unfoldly helps you find files you can't remember by describing what you're looking for, making it perfect for messy folders and personal files. It's local-first and private, ensuring your files stay on your device and don't require an account.

Overview: This project implements the core ideas from the PagedAttention paper, addressing memory management challenges in LLM inference. It introduces block-based KV cache, non-contiguous memory, copy-on-write, swap and recompute strategies, and reduced memory fragmentation. Key features: - Correctness verified: Outputs match vanilla attention within 1e-5 tolerance - Memory efficient: 60-80% memory savings vs naive allocation - Parallel sampling: Shares prompt blocks across samples - Beam search: COW semantics for efficient forking - Flexible swapping: Simulates CPU-GPU transfers - Comprehensive tests: Unit tests and benchmarks included Tech stack: Python Why it matters: This project offers an efficient solution for memory management in LLM inference, significantly improving LLM serving throughput and reducing memory fragmentation.

### Overview: nano-vectordb-rs is a simple, easy-to-hack vector database implementation written in Rust. It provides fast cosine similarity searches using Rayon parallelism and is embedded with compact serialization for efficient storage. ### Key Features: - **Fast Cosine Similarity Searches**: Utilizes Rayon parallelism for quick and efficient searches. - **Simple API Surface**: Easy to use with minimal abstraction. - **Embedded Persistence**: Compact serialization for embedded persistence. - **No Abstractions**: Simple, clean implementation for easy hacking. ### Tech Stack: Rust ### Why It Matters: nano-vectordb-rs is a Rust port of the popular nano-vectordb, making it a perfect choice for Rust ML pipelines, prototyping semantic search systems, and educational use cases. Its lightweight nature and clean, hackable implementation make it an excellent choice for developers looking to quickly integrate vector database functionality into their projects.

### Overview MovieMatch is a movie recommendation system that uses a hybrid ranking approach to suggest films similar to your preferred ones. It provides a ranked top-10 list, allows sorting by various criteria, and offers AI insights via a local Ollama model for enhanced analysis. ### Key Features - **Hybrid Ranking**: Combines semantic similarity, rating, and popularity to rank recommendations. - **AI Insights**: Detailed analysis comparing your seed movie to each pick, including plot summary, audience fit, and discussion questions. - **Rich Movie Cards**: Instant access to genres, cast, director, rating, match percentage, and release year. - **Sortable Results**: Reorder recommendations by best match, rating, popularity, or release year. ### Tech Stack - **Frontend**: Jupyter Notebook (via Next.js) - **Backend**: FastAPI with FAISS and Ollama integration ### Why It Matters MovieMatch simplifies the process of discovering new movies that align with your tastes, making it easier to explore and enjoy a wide range of films. The AI insights provided by Ollama enhance the user experience by offering deeper analysis and context around each recommendation. ### Repository - **Repository**: VivekaAryan/Movie_recommendation_system - **Stars**: 9 - **Description**: Helps discover films similar to ones you love. Search by title, get a ranked top-10 list, sort results, and optionally generate AI-powered insights for each pick via a local Ollama model. ### Project Structure - **Architecture**: Next.js frontend and Fast

Overview: VectorLite is a Rust implementation of the SQLite database optimized for vector databases, providing compact size, zero-copy de/serialization, excellent search performance, and minimal dependencies. Key features: - Compact size - Zero-copy de/serialization from/to disk - Excellent search performance - WASM support - Minimal dependencies Tech stack: Rust Why it matters: VectorLite offers a highly efficient and lightweight alternative to traditional vector databases, making it ideal for applications requiring fast and scalable search capabilities.

### Overview: Costume Director Vault Starter 是一个服装视觉知识库，旨在整理服装结构、材质、Prompt与角色造型研究笔记。它通过结构、材质、表面、轮廓和造型的详细描述，帮助设计师和研究者更好地理解和复现服装设计。 ### Key features: - 结构卡：整理服装结构单元，如领型、袖型、门襟等。 - 材质卡：描述纤维、组织、表面、光泽、垂坠等材质特征。 - 单品卡：从单品进入版型、结构和穿着方式。 - 风格卡：涵盖历史风格、当代风格等。 - 造型卡：完整人物造型、场景、角色、鞋履、配饰与跨镜连续性。 - 视觉提示词：复制到生图工作流，或作为视觉沟通文本。 - 参考图状态：区分“有图”“有许可”“已关联”“已在 Obsidian 显示”。 ### Tech stack: Obsidian 卡片系统，Markdown 语法，DASH Skill 插件。 ### Why it matters: 服装设计中，最难的往往是把名词还原成可辨认的视觉事实。这个仓库试图把这些看不见的知识整理成一张可以反复查阅、组合和验证的工作台，帮助设计师和研究者更好地理解和复现服装设计。 ### Repository: - **Repository**: YanKaFei/Costume-Director-Vault-Starter - **Stars**: 9 - **Description**: 服装视觉知识库，结构

Overview: This project is an unofficial fork of Qwen3.8-Flash-Next EXL3, optimized for Flash Next and designed for high-performance tasks. It features a 3.05bpw prefill speed of 1777 t/s, making it suitable for tasks requiring prompt-lookup drafter capabilities. The project includes 262k context, making it ideal for applications that require efficient use of memory and processing power. Key features: - Flash Next Optimization - Prompt-lookup Drafter - Native Vision/Video Support - 262k Context Tech stack: Python Why it matters: This project is designed to enhance performance and efficiency in Flash Next tasks, making it a valuable tool for developers looking to optimize their applications.

Overview: Svoi is a self-hosted, open-source personal AI assistant that streams chat with agentic tools, long-term memory, and a RAG knowledge base. It uses Go, Vue 3, PostgreSQL, and OpenAI-compatible LLMs. Key features: - Streaming chat with agentic tools: web_search, browse, generate_image. - Long-term memory: facts with pgvector embeddings, extraction, summaries, and a forget-list. - Reverse proxy: nginx for serving the SPA and proxying API requests. - PostgreSQL with pgvector for database storage. Tech stack: Go 1.25, standard `net/http` (Go 1.22+ `ServeMux`), pgx/v5 + sqlx, hand-written SQL. Why it matters: Svoi offers a self-hosted, open-source solution for personal AI assistants, providing a streaming chat experience with agentic tools, long-term memory, and a RAG knowledge base. The use of PostgreSQL and pgvector ensures robust data management and efficient querying. The reverse proxy setup with nginx simplifies the deployment process, making it accessible for a wide range of users.

### Overview: This repository contains a PyTorch implementation of **MVC-ZigAL**, as presented in our paper [*Refining Few-Step Text-to-Multiview Diffusion via Reinforcement Learning*](https://arxiv.org/abs/2505.20107). ### Key Features: - MVC-ZigAL is a tailored Reinforcement Learning finetuning framework designed specifically for few-step Text-to-Multiview Diffusion models. - It addresses the challenges of few-step T2MV diffusion models by proposing a new MDP formulation that jointly models all generated views and assesses their collective quality via a joint-view reward. - A novel advantage learning strategy is introduced, which exploits the performance gains of a self-refinement sampling scheme over standard sampling, yielding stronger learning signals for effective RL finetuning. - A unified RL framework is extended with a Lagrangian dual formulation for multiview-constrained optimization, balancing single-view and joint-view objectives through adaptive primal-dual updates under a self-paced threshold curriculum. ### Tech Stack: - Python ### Why it Matters: MVC-ZigAL offers a robust and balanced Reinforcement Learning finetuning approach for few-step T2MV diffusion models, significantly improving both per-view fidelity and cross-view consistency compared to existing methods.

Overview: ModelDock is an open-source ML model registry and inference platform designed for versioned artifact management, deployment, inference, benchmarking, and observability. Key features: - Versioned model management - Artifact integrity verification - Multiple runtimes for Python, JSON, and scikit-learn - Explicit deployment lifecycle with audit history and evaluation gates - Deployment audit trail - Inference API with version-aware requests - Asynchronous inference jobs with durable status - Runtime caching with safe artifact replacement invalidation - Restricted Python execution with import, dunder, and unsafe builtin checks - Authentication with configurable API key protection - Metrics and inference history for operational visibility - Dockerized development with PostgreSQL and Redis - Next.js dashboard for models, inference, history, and monitoring - Automated CI/CD for tests, compilation, builds, migrations, security scanning, container validation, and release images Tech stack: Python Why it matters: ModelDock provides a comprehensive solution for ML infrastructure, ensuring model versioning, artifact integrity, deployment, inference, and observability, making it a valuable tool for developers and data scientists. Repository: aawhan0/ModelDock GitHub stars: 9 GitHub description: Open-source ML model registry and inference platform with versioned artifacts, predictions, metrics, and observability.

Overview: This beginner-to-master roadmap is designed to help modern AI engineers gain a comprehensive understanding of AI and ML fundamentals, deep learning, language models, retrieval-Augmented Generation (RAG), agents, model infrastructure, optimization techniques, hardware acceleration, AI security, blockchain, Zero-Knowledge Merkle Ladder (ZKML), and culminate in capstone mastery. Key features: - AI and ML fundamentals - Deep learning - Language models (LLMs) - Retrieval-Augmented Generation (RAG) - Agents - Model infrastructure - Optimization techniques - Hardware acceleration - AI security - Blockchain - Zero-Knowledge Merkle Ladder (ZKML) - Capstone mastery Tech stack: Python Why it matters: This roadmap provides a structured approach to mastering modern AI and ML, covering essential topics and practical applications, making it ideal for aspiring AI engineers looking to build a strong foundation and advance their skills.

### Overview Solar filaments are dark, elongated structures of cool plasma in the solar atmosphere. Detecting them automatically is crucial for space weather research. This project uses deep learning semantic segmentation to predict binary masks of filaments in full-disk H-alpha images of the Sun. It compares a U-Net trained from scratch with a U-Net fine-tuned with EfficientNetB0 transfer learning. ### Key Features - **Baseline U-Net** trained from scratch - **EfficientNetB0 Transfer Learning** with fine-tuning of top encoder layers - **Fine-tuning of loss weights and prediction threshold** - **Dice coefficient evaluation** for model comparison - **Kaggle submission pipeline** at higher resolution (1024 × 1024) with per-filament instance segmentation ### Tech Stack - **TensorFlow/Keras** - **EfficientNetB0** for transfer learning - **U-Net** architecture for semantic segmentation ### Why It Matters - **Automated Filament Detection** is essential for space weather research, enabling better prediction of solar eruptions. - **Comparative Analysis** of multiple models helps in identifying the most effective approach for automatic filament detection. - **Transfer Learning** leverages pre-trained models to save training time and improve performance. ### Repository - **GitHub Repository**: ali-119/Solar-Filament-Segmentation - **Stars**: 5 - **Description**: Deep learning semantic segmentation project for solar filament detection using TensorFlow/Keras, benchmarking custom CNN

### Overview: 文本到图像合成技术在PyTorch中实现，通过GAN和LSTM技术，有效防止了判别器的退化现象，提升了生成模型的效果。 ### Key features: - 文本到图像任务，使用GAN和LSTM技术。 - 词信息和句子信息共同指导生成器训练。 - 实现了Inception Score指标，训练参数缩减了一个数量级。 - 通过对比不同粒度的语言信息和图像信息的结合顺序，证明了其对GAN训练的影响。 ### Tech stack: Python ### Why it matters: 该模型不仅在文本到图像生成方面取得了显著的进展，而且在训练参数上也实现了巨大优化，具有重要的应用价值和研究意义。

Overview: This project involves fine-tuning the Llama2-7B model from scratch using LoRA and QLoRA techniques to generate structured SOAP notes from patient-doctor conversations. The dataset used for training comprises transcribed medical dialogues formatted into SOAP notes. Key features: - Fine-tuning of the Llama2-7B model - Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA) for efficient training - Preprocessing of custom dataset into the required format Tech stack: Python 3.11 or higher, PyTorch 1.10.0 or higher, CUDA 10.2 or higher Why it matters: This project automates the generation of structured SOAP notes from patient-doctor conversations, improving efficiency and accuracy in clinical documentation.

Overview: This repository contains the code and data for a research paper titled "LLM as a Broken Telephone: Iterative Generation Distorts Information." The project investigates how Large Language Models (LLMs) distort information when used iteratively, mirroring the "broken telephone" game where messages degrade through repeated transmission. Key features: - Iterative generation of text using Hugging Face Transformers models. - Simulating the "broken telephone" effect in machine translation and rephrasing tasks. - Quantifying distortion using BLEU, ROUGE, CHR-F, METEOR, BERTScore, and FActScore. Tech stack: Python Why it matters: This research provides valuable insights into the limitations of LLMs and highlights the importance of understanding how these models process and transmit information, particularly in scenarios where information can be easily distorted.

Overview: PDFGenie is a PDF-based chatbot that leverages vector embeddings and AI chat models to answer questions from PDF documents. Key features: - Upload PDFs and generate vector embeddings for fast semantic search. - Ask questions and get answers from PDF content using AI-powered chat. - Seamlessly integrates Pinecone vector database for efficient search. - Easy configuration using environment variables. Tech stack: Node.js / Express.js, Gemini API, GROQ API, Pinecone. Why it matters: PDFGenie simplifies the process of extracting information from PDFs, making it accessible and searchable using AI-powered chatbots, enhancing productivity and knowledge management.

### Overview: A fine-tuned SpeechT5 model for Urdu text-to-speech generation with voice cloning capabilities. Supports both Urdu and Roman Urdu scripts, allowing speaker selection for personalized speech synthesis. ### Key Features: - 🗣️ **Urdu TTS**: High-quality text-to-speech synthesis for Urdu language - 🔊 **Voice Cloning**: Generate speech in the style of specific speakers (including Zia Mohiuddin's voice) - 🌐 **Dual Script Support**: Works with both Urdu (نثر) and Roman Urdu (Urdu written in Latin script) - 🎛️ **Speaker Selection**: Choose between different voice profiles - 🚀 **FastAPI Demo**: Interactive web interface for testing the model ### Tech Stack: - **Model**: SpeechT5 (fine-tuned) - **Backend**: FastAPI - **Frontend**: HTML/CSS/JavaScript - **Audio Processing**: Librosa, SoundFile - **ML Framework**: PyTorch, Transformers ### Why it Matters: This model offers expressive, speaker-specific synthesis with a FastAPI demo for easy testing, making it suitable for applications requiring high-quality speech output in Urdu.

### Overview: Narwhal is an adaptive disaggregated inference framework designed for hot-swapping prefill and decode roles on a single GPU or multi-node deployment, ensuring seamless performance and scalability. ### Key features: - **Role Hot-Swap**: Automatically reassigns prefill and decode roles across a fixed GPU fleet, using NIXL key-value (KV) transfers. - **Serving**: Serves completion and chat requests with latency-aware admission, supporting both streaming and buffered requests. - **Fault Tolerance**: Fails over to a warm-standby router and re-admits engines against their live process generation. - **Measurement**: Profiles engines, runs ordered benchmark points, and keeps the evidence from each point. - **Observability**: Exports router and engine metrics to Prometheus and a provisioned Grafana dashboard. - **Operator Tooling**: Validates fleet files offline and collects private diagnostic bundles. ### Tech stack: Python 3.11 through 3.13 ### Why it matters: Narwhal simplifies the management of complex inference tasks by dynamically reassigning roles, ensuring optimal performance and scalability across different deployment scenarios.

### Overview: This project is a Question Answering application utilizing Large Language Models (LLMs) and Amazon OpenSearch Serverless Service. It employs the RAG (Retrieval Augmented Generation) approach to retrieve and bundle relevant information from an enterprise knowledge base or content, then sends it to the LLM for a GenAI response. Amazon OpenSearch Serverless is used for storing embeddings, enhancing the accuracy of the LLM by selecting the most pertinent passages. ### Key Features: - Amazon OpenSearch Serverless Service for storing and retrieving embeddings. - RAG (Retrieval Augmented Generation) approach for context enrichment. - SageMaker Endpoint for text generation and embedding generation. - Amazon SageMaker Studio for developing and deploying the application. ### Tech Stack: Python ### Why it Matters: Amazon OpenSearch Serverless Service offers a scalable and cost-effective way to store and retrieve embeddings, enhancing the accuracy of the LLM by selecting the most relevant passages from the enterprise knowledge base. This project demonstrates how to effectively integrate large language models with Amazon OpenSearch Serverless to improve the quality of GenAI responses, making it a valuable resource for developers looking to build powerful question answering applications.

### Overview: Grok Imagine 1.5 Edit prompts have been tested with 8 recipes, showcasing effective image prompt strategies for various subjects and lighting setups. ### Key Features: - **Model ID:** `grok-imagine-1.5-edit-apimart` - **Aspect Ratios:** 1:1, 3:4, 4:3, 8:5, 16:9, 2:3, 1:1.5, 2:1 - **Cost Per Image:** $0.015 ### Tech Stack: Python ### Why It Matters: - **Efficient Pricing:** One OpenAI-compatible endpoint at $0.015 per image, no subscription required. - **Pay-As-You-Go:** No monthly plan, pay only for the calls. - **Immediate Pricing:** The cost is reported in the response, ensuring accurate billing. - **Asynchronous Calls:** Tasks can be batched and retried, making integration straightforward. ### Repository: b5ap9jopbp1b6/grok-imagine-1-5-edit-prompts ### Primary Language: Python ### GitHub Stars: 9 ### GitHub Description: Grok Imagine 1.5 Edit prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`grok-imagine-1.5-edit-apimart`) and the cost per image.