Projects

Overview: DocSense is an interactive web application designed for comparing and analyzing PDF documents using cosine similarity scans and a chatbot interface. Key features: - Document-Level Cosine Similarity Scan - Sentence-Level Cosine Similarity Scan - Preprocessed Sentence-Level Cosine Similarity Scan - ChatBot for interacting with PDF content Tech stack: Streamlit, FastAPI, LangChain Why it matters: DocSense simplifies the process of comparing and analyzing PDF documents by providing a user-friendly interface and advanced text analysis capabilities, making it valuable for researchers, educators, and anyone needing to compare document content efficiently.

Overview: ResFit is an implementation of Residual Off-Policy RL for Finetuning Behavior Cloning Policies, focusing on improving the performance of behavior cloning policies through residual off-policy learning. Key features: - Residual off-policy reinforcement learning algorithm - Finetuning behavior cloning policies - Implementation in PyTorch Tech stack: Python Why it matters: ResFit offers a novel approach to behavior cloning by leveraging residual off-policy learning, potentially improving the accuracy and efficiency of policy finetuning compared to traditional methods. Repository: lucidrains/resfit-pytorch Primary language: Python GitHub stars: 18

Overview: This project is a production-oriented FastAPI starter for an Agentic RAG system, focusing on authenticated, document-scoped, and evaluation-driven retrieval augmented generation. It includes features like user authentication, document ownership, semantic cache, and a simple built-in UI for managing evaluations. Key features: - Authenticated users and owned documents - Document-scoped retrieval to prevent cross-document leakage - Text and PDF ingestion - Pluggable embeddings and reranking - Semantic answer cache backed by Postgres + pgvector - Evaluation pipeline with run history, case inspection, and rerun support - Simple built-in UI for asking questions and running evaluations Tech stack: FastAPI, SQLAlchemy, Alembic, PostgreSQL, pgvector, Chroma, OpenAI, Hugging Face, Cohere, JWT, Docker, Docker Compose, Jinja templates, and FastAPI Users. Why it matters: This project addresses the need for a production-ready, modular Agentic RAG system that supports authentication, document-scoped retrieval, and evaluation-driven workflows, making it suitable for various applications requiring a clean and efficient RAG implementation.

Overview: Vectorstores is a framework designed to simplify the integration of vector databases into AI applications, making it easier to connect data to your AI application and query it later. Key features: - Unified interface for vector database connections - Support for ingestion of data from various sources - Lightweight implementation - Compatibility with multiple JavaScript environments Tech stack: TypeScript Why it matters: Vectorstores simplifies the integration of vector databases into AI applications, making it easier to connect data to your AI application and query it later. It is lightweight and compatible with multiple JavaScript environments, making it a valuable tool for developers looking to enhance their AI applications.

### Overview: Totem is a file-based toolkit that stores project rules, lessons, and context directly within the repository. It leverages plain markdown lessons, a local knowledge index derived from these lessons, and a zero-LLM lint engine for deterministic and offline linting. ### Key Features: - **Plain Markdown Lessons**: Store project-specific rules and lessons. - **Local Knowledge Index**: A local, dynamic index derived from the repository's files. - **Deterministic Linting**: A zero-LLM lint engine that enforces lint rules. - **Queryable Knowledge Index**: A local, dynamic index derived from the repository's files. ### Tech Stack: TypeScript ### Why It Matters: By keeping project rules and lessons in the repository alongside the code, Totem ensures that the team's learnings persist across sessions, eliminating the need for constant re-explanation and architectural mistakes. This approach simplifies the workflow and reduces the friction of maintaining architectural integrity, making the team's work more efficient and reliable. --- ### Repository: mmnto-ai/totem ### Primary Language: TypeScript ### GitHub Stars: 18 ### GitHub Description: Rules you can enforce, state you can derive, context you can query: plain files in your repo, under whichever coding agent you use.

### Overview: Weaviate Studio is a powerful VS Code extension designed for managing and querying Weaviate vector databases. It offers an intuitive GraphQL interface, supporting both self-hosted and cloud instances. ### Key Features: - **Unified Experience:** Direct management and querying of Weaviate databases within VS Code. - **Works Everywhere:** Connects to local, on-prem, or cloud Weaviate instances. - **Visual Data Exploration:** Interactive data browser with filtering, vector search, and export. - **Intelligent GraphQL Editor:** Auto-complete, schema-aware templates, and error highlighting. - **Secure & Productive:** Secure credential storage, hot reload, and type-safe development. ### Tech Stack: TypeScript ### Why it Matters: Weaviate Studio simplifies the management and querying of Weaviate vector databases, making it a valuable tool for developers working with large-scale vector data.

Overview: This roadmap is designed for software engineers with Python and basic ML knowledge who aspire to break into AI/LLM roles. It provides a structured path with free and paid courses, hands-on projects, and interview prep to help candidates land AI/LLM engineering jobs. Key features: - Free and paid courses - Hands-on projects - Interview prep - Covers Transformers, RAG, fine-tuning, agents, LangChain, MLOps, and deployment Tech stack: Python, ML basics Why it matters: This roadmap offers a structured, no-fluff approach to becoming an AI/LLM engineer, focusing on the skills and projects that top companies are looking for. Repository: musamaanjum/ai-engineer-roadmap GitHub stars: 18 GitHub description: A complete roadmap to become an AI/LLM Engineer in 2026. Free & paid courses, 5 portfolio projects, interview prep, and resources. For software engineers with Python & ML basics who want to break into AI. Covers Transformers, RAG, fine-tuning, agents, LangChain, MLOps, and deployment.

### Overview: This repository, **NekoAI-agents**, serves as a comprehensive guide and library for developers interested in AI agents and related technologies. It provides foundational knowledge, hands-on examples, and advanced patterns for creating intelligent agents using Python, GCP, PostgreSQL, and other tools. ### Key Features: - **Foundation**: Covers core concepts and building blocks for AI agents, including understanding what agents are, implementing memory systems, MCP integrations, and vector databases. - **Claude Projects**: Demonstrates real-world AI agent implementations using Claude Code, showcasing advanced patterns like custom agents, skills, and domain-specific workflows. - **Books and References**: Curated learning resources for deepening understanding of AI agents, vector databases, memory management, and embeddings. ### Tech Stack: - Python - GCP (Google Cloud Platform) - PostgreSQL - OpenAI API - Gemini Models - Langchain Package ### Why It Matters: This repository is essential for developers looking to build intelligent agents and understand the foundational concepts and advanced patterns required for creating effective AI solutions. The hands-on examples and curated learning resources make it a valuable resource for both beginners and experienced developers alike. ### Repository Information: - Repository: nekozzz94/NekoAI-agents - GitHub Stars: 18 - GitHub Description: AI agent code examples - Primary Language: Python

### Overview: XTLLM, an experimental text-only Vulkan inference engine, is designed to run very large mixture-of-experts models on consumer AMD GPUs. It leverages a model-specific working set in VRAM, a budgeted expert tier in system RAM, and explicit host-memory context for long-context execution. ### Key Features: - **Native Vulkan Inference**: Runs inference using Vulkan, leveraging the GPU's parallel processing capabilities. - **Model-Specific Working Set**: Keeps a model-specific working set in VRAM to optimize performance. - **Budgeted Expert Tier**: Uses system RAM to manage an expert tier, reducing memory usage. - **Explicit Host Memory Context**: Host memory is explicitly accounted for in the GPU path, ensuring efficient use of resources. ### Tech Stack: - **C++**: The primary programming language used for the implementation. ### Why It Matters: XTLLM is designed to run very large models on consumer GPUs, providing a practical solution for inference tasks. It offers a balance between model size and performance, making it suitable for applications that require high throughput and low latency. ### Repository: opktunme/XTLLM - **GitHub Stars**: 18 - **GitHub Description**: Native Vulkan inference for giant MoE models on consumer AMD GPUs ### Primary Language: C++ - **GitHub Description**: Native Vulkan inference for giant MoE models on consumer AMD GPUs ### Why This Exists: - **Experimental Design**: Designed to run very large models on consumer GPUs, providing a practical solution for inference tasks. - **GPU Compatibility**: Hardware-validated on the

### Overview: Open Translator is a versatile tool designed for voice-to-voice and speech-to-text translation, featuring support for over 17 languages. It offers both online and local modes, including voice cloning capabilities, making it suitable for various translation needs. ### Key Features: - **Translation from and to 17 Languages**: Supports English, Spanish, French, German, Dutch, Japanese, Korean, Turkish, Arabic, Russian, Hebrew, Hindi, Italian, Portuguese, Chinese, Czech, and Hungarian. - **Online Mode**: Translates speech to text and text to speech using Google Translate. - **Local Mode**: Utilizes local language models for speech-to-speech translation. - **Audio File Conversion and Extraction**: Converts audio files to MP3, extracts audio from videos, and replaces audio in videos. - **YouTube Downloader and Video Text Adder**: Provides functionality to download YouTube videos and add text to videos. - **Voice Recorder and PyTranscriber**: Offers a voice recorder and a shortcut for PyTranscriber. - **Demo Video**: Demonstrates the tool's features through a series of pre-recorded examples. ### Tech Stack: Python ### Why It Matters: Open Translator is a powerful tool for real-time translation and voice processing, making it invaluable for language learners, remote communication, and content localization. Its ability to translate speech to text and vice versa, combined with its support for multiple languages, makes it a versatile solution for various translation needs.

### Overview: nanaban is a command-line interface (CLI) tool that generates AI-generated images directly from terminal prompts. It utilizes GPT Image 2, Nano Banana, and GPT-5 Image models, offering a seamless workflow for both human users and LLM agents. ### Key Features: - **One Command, One Image**: Generate an image with a single command. - **Free GPT Image 2 for ChatGPT Plus/Pro Subscribers**: Utilizes OAuth token from ChatGPT Plus/Pro subscriptions for free image generation. - **Three Model Families**: Supports GPT Image 2, Nano Banana, and GPT-5 Image. - **Auto-Names Files**: Automatically generates file names based on the prompt. - **Built for Scripts**: Outputs to stdout, suitable for scripts and CI pipelines. - **Built for LLM Agents**: Compatible with `--json` for structured output. ### Tech Stack: TypeScript ### Why It Matters: nanaban simplifies the process of generating AI images, making it accessible and efficient for both humans and LLM agents. It eliminates the need for browser tabs, queues, or API key management, ensuring a seamless and cost-effective workflow.

Overview: RadCrew Unveiled is a monorepo consisting of a Vite React frontend and a FastAPI backend API. It includes a LLM-powered chatbot that interacts with a static site and GitHub Markdown for content retrieval. Key features: - A monorepo structure for seamless development and deployment. - A chatbot that leverages a static site and GitHub Markdown for content retrieval. - A backend API using Python and FastAPI for handling requests and providing responses. Tech stack: TypeScript Why it matters: This project showcases the integration of a frontend and backend using modern technologies, demonstrating how to build a system that can interact with static content and external data sources like GitHub, making it useful for developers looking to create similar applications.

Overview: This repository documents the setup and performance of a Flash Next model running on two RTX 3090 GPUs, with benchmarks and details on the setup and configuration. Key features: - Flash Next model running on two RTX 3090 GPUs - Benchmarks of writing and reading prompts - Full set of benchmarks in `BENCHMARKS.md` - Configuration options for different prompt lengths Tech stack: Python Why it matters: This repository provides insights into the performance and capabilities of the Flash Next model, focusing on its ability to handle large text inputs and generate responses efficiently. Repository: `ruashots/flashnext-2x3090` GitHub stars: 17

Overview: This project allows users to interact with PDF documents through a conversational interface, leveraging RAG models and Anthropic Claude 3 for intelligent responses. Key features: - PDF Upload - Text Extraction - Text Chunking - Vector Store Creation - Conversational Interface Tech stack: Python, Streamlit, PyPDF2, Langchain, FAISS, Anthropic. Why it matters: This project provides a practical application of RAG models and LLMs, enhancing the accessibility and efficiency of information retrieval from PDF documents.

### Overview: PACS: A Dataset for Physical Audiovisual Common-Sense Reasoning is a repository containing data and code for the paper "PACS: A Dataset for Physical Audiovisual CommonSense Reasoning" presented at ECCV 2022. This dataset aims to facilitate research in physical audiovisual common-sense reasoning, providing a comprehensive resource for understanding and improving common-sense reasoning abilities. ### Key Features: - **Data and Code**: Contains both the dataset and the accompanying code for the research paper. - **Model Baselines**: Provides benchmark models for comparison, including Fusion, CLIP, UNITER, Merlot Reserve, and Majority. - **Citation**: The repository includes a citation for the paper, encouraging users to cite it in their own work. ### Tech Stack: Python ### Why It Matters: This dataset and accompanying code are crucial for advancing research in physical audiovisual common-sense reasoning, offering a valuable resource for researchers and practitioners looking to improve their understanding and implementation of common-sense reasoning in physical and audiovisual contexts.

### Overview: This repository contains an implementation of the Calibrating LLMs with Information-Theoretic Evidential Deep Learning (IB-EDL) method, focusing on fine-tuning Llama3-8B on the OBQA dataset and evaluating the performance on the CSQA dataset for OOD detection. The post-hoc calibration technique is also introduced for further performance enhancement. ### Key features: - Implementation of IB-EDL for fine-tuning LLMs on multiple-choice QA tasks. - Training and evaluation on the OBQA and CSQA datasets. - Post-hoc calibration technique for improving IB-EDL performance. - Adaptation to generative tasks for open-ended generation. ### Tech stack: Python ### Why it matters: This repository provides a practical implementation of IB-EDL, demonstrating its effectiveness in calibrating LLMs and evaluating their performance on diverse datasets. The post-hoc calibration technique offers an additional layer of performance enhancement, making it valuable for researchers and practitioners looking to improve LLMs' reliability and generalization capabilities.

### Overview: Helios is a small, readable PyTorch inference engine for running Qwen3-4B on a local NVIDIA GPU. It implements the model, weight loading, tokenization, KV caching, generation, and HTTP serving path directly in PyTorch. ### Key features: - A native PyTorch implementation of Qwen3-4B - Grouped-query attention, rotary position embeddings, RMS normalization, and SwiGLU feed-forward layers - Native paged attention through PyTorch `varlen_attn`, using a shared pool of 256-token KV pages - Hugging Face `tokenizer.json` and safetensor weight loading - Packed mixed batches that can run decode tokens and prefill chunks in one model forward - Continuous batching of concurrent requests - Per-request page tables and a persistent prefix cache of 256-token prompt blocks - GPU-memory admission checks and a shared budget for active KV and cached prefixes - A non-streaming OpenAI-style chat completions endpoint - A concurrent HTTP benchmark with per-request and aggregate timing data ### Tech stack: Python ### Why it matters: Helios is an inference-engineering learning project, not a production serving system. The code favors mechanisms that are easy to inspect and change, not a broad feature set.

### Overview: `ferrovec` is an in-browser vector store that persists to disk and remains consistent across tabs, ensuring semantic search remains intact even after a reload or tab switch. It uses a hand-rolled Hierarchical Navigable Small World (HNSW) graph engine, providing a durable and cross-tab consistent solution. ### Key Features: - **Durable by Default**: The index persists to OPFS and rehydrates on `open()`. - **Cross-tab Consistent**: Single-writer leader election (Web Locks + BroadcastChannel) ensures consistency across multiple tabs. - **Incremental**: Supports upsert-style inserts, tombstoning removals, and in-place `compact()` operations. - **Real HNSW, in Rust**: Hand-rolled HNSW graph with Pinecone, Weaviate, and Qdrant algorithm. - **Featherweight & Shim-Free**: Dependencies are `serde` and `postcard`, with a 77 KB (35 KB gzipped) core. - **SIMD-accelerated**: Distance calculations are SIMD-accelerated with a scalar fallback, and `#![deny(unsafe_code)]` everywhere outside the audited kernel. ### Tech Stack: Rust → WASM ### Why It Matters: - **Durable and Shared**: The HNSW graph lives on disk, ensuring semantic search is consistent across tabs. - **Incremental Updates**: Supports upsert-style inserts, tombstoning removals, and in-place `compact()` operations. - **Cross-Tab Consistency**:

Overview: PAPERCLIP is a method that aligns astronomical observations with natural language using a neural network model fine-tuned from a pre-trained Contrastive Language-Image Pre-training (CLIP) model. The model is trained on successful observing proposal abstracts and downstream observations, optionally summarized using large language models (LLMs). The study demonstrates the potential of using generalist foundation models for interacting with astronomical data through meaningful joint representations. Key features: - Fine-tuned CLIP model - Association of astronomical observations with natural language - Use of successful observing proposal abstracts for training - Optional summary using large language models (LLMs) - Evaluation of image retrieval and description retrieval Tech stack: Jax Why it matters: PAPERCLIP shows the potential for using generalist foundation models rather than task-specific models for interacting with astronomical data by leveraging text as an interface, demonstrating the ability to effectively represent and retrieve astronomical observations through natural language queries.

### Overview: DeepUnity is an open-source deep learning library designed specifically for Unity environments, providing advanced tensor computation and reinforcement learning tools. It supports GPU acceleration and offers state-of-the-art algorithms for intelligent agent training. ### Key Features: - **Tensor Computation with GPU Acceleration**: Enables efficient tensor operations on GPUs for faster performance. - **Deep Neural Networks**: Supports complex neural network architectures for advanced machine learning tasks. - **Reinforcement Learning Tools**: Facilitates training of intelligent agents using reinforcement learning algorithms. ### Tech Stack: C# ### Why It Matters: DeepUnity simplifies the integration of deep learning into Unity environments, making it easier to develop complex AI applications without the need for specialized knowledge in deep learning frameworks. ### Repository: smtmRadu/DeepUnity ### GitHub Stars: 18