Projects

Overview: Fortémi is an intelligent database designed for AI-ready applications, providing a normalized data schema and tooling for converting messy organizational data into searchable and linkable structures. Built in Rust, it supports hybrid semantic search, multi-provider LLM inference, and multimodal ingestion, making it suitable for various use cases such as knowledge management, RAG, and team-documentation workflows. Key features: - **Search That Misses the Point**: Traditional search relies on keywords, which can miss the actual content of the data. Fortémi uses semantic search to find relevant records even when terminology differs. - **MCP Server**: A server-side component for multi-provider LLM inference, supporting Ollama, OpenAI, OpenRouter, and llama.cpp. - **Multi-Provider LLM Inference**: Integrates multiple providers for seamless inference, ensuring diverse and accurate responses. - **Multimodal Ingestion**: Capable of ingesting data from various sources including vision, audio, and text, making it versatile for different types of information. - **Knowledge Graph**: A knowledge graph to connect and understand the relationships between different pieces of information, enhancing the overall data structure. Tech stack: Rust + PostgreSQL + MCP server (Node.js). Why it matters: Fortémi provides a robust and privacy-first solution for AI-ready applications, ensuring that data is searchable, linkable, and provenance-aware. Its ability to handle diverse data sources and provide intelligent search capabilities makes it a valuable tool for organizations looking to improve their data management and AI integration. Repository: Fortemi/fortemi

Overview: OllaLab-Lean is a lean stack designed to help both novice and experienced developers rapidly set up and begin working on LLM-based projects, accelerating local Large Language Model (LLM)-based research and development. Key features: - Pre-made prompt templates and applications to accelerate research and developments. - Ollama for managing local openweight LLMs. - LangChain for orchestrating LLM pipelines, allowing users to seamlessly connect, manage, and optimize their workflows. - Streamlit server to locally host dynamic LLM-based web applications. - Jupyter Lab server as the integrated development environment (IDE), providing users with an interactive space to write, test, and iterate code efficiently. - Neo4J vector database supporting retrieval-augmented generation (RAG) tasks. - Data analysis, AI, ML tools such as: DuckDB, AutoGluon, AutoViz, GenSim, etc. - Monitoring and Logging tools such as: Elastic Search, Kibana, Grafana, Prometheus. Tech stack: Python, Jupyter Notebook, Docker, Podman, Neo4J. Why it matters: OllaLab-Lean streamlines the setup and workflow for developers, allowing them to focus on research and development rather than on the initial configuration of complex environments. This makes it particularly useful for both beginners and experienced developers looking to quickly get started with LLM-based projects.

Overview: This project provides a memory-efficient implementation of language models optimized for training on laptops, leveraging LoRA fine-tuning techniques. Key features: - Supports multiple models like Qwen/Qwen2.5-0.5B-Instruct and DeepSeek-R1-Distill-Qwen-1.5B. - Implements SLLM (Super Lazy Language Model) for efficient memory usage. - Bundles scaled matmul operations for parallel execution across CPU cores. - Optimized for reduced compute intensity and GPU utilization. Tech stack: Python Why it matters: This library simplifies the training process for large language models on low-power devices, reducing memory and computational requirements. Repository: HenryNdubuaku/super-lazy-autograd GitHub stars: 30 GitHub description: Hand-derived memory-efficient VJPs for tuning LLMs on laptops.

### Overview: DFRot aims to address the limitations of randomized Hadamard transforms in large language models (LLMs) by introducing refined orthogonal transforms. This project explores the concept of outlier-free and massive activation-free rotation, providing a significant improvement in performance. ### Key Features: - **Outlier-Free Rotation**: Achieves rotation without introducing outliers. - **Massive Activation-Free Rotation**: Ensures massive activation without degradation. - **Refined Rotation Techniques**: Utilizes orthogonal transforms for better performance. ### Tech Stack: Python ### Why it Matters: DFRot offers a more effective and efficient approach to rotation in LLMs, leading to improved model performance and reduced computational overhead.

Overview: Arvutaja is an Android app designed for voice actions, translating spoken utterances into formal expressions and commands in Estonian and English. Currently supported languages and expressions include measurement unit conversions, currency conversion, arithmetical expressions, alarm clock commands, phone number queries, and address and weather queries. Key features: - Voice actions app for Estonian and English - Grammar-based speech recognition using Action-grammar - Persistent query history list - Support for vague/ambiguous queries - Third-party app interaction for map, currency conversion, and second opinion - Open and modular design with human-readable grammar - Uses open source grammars and avoids dependency on speech recognition servers Tech stack: Java Why it matters: Arvutaja offers a unique approach to voice actions by being largely defined by a human-readable multilingual grammar, providing an open and modular solution that is not dependent on specific speech recognition services. Repository: Kaljurand/Arvutaja GitHub stars: 30 GitHub description: An Android app for voice actions in Estonian and English

Overview: This is a Dockerfile designed to run a DeepSpeech server, providing a scalable and efficient solution for speech recognition tasks. Key features: - Run DeepSpeech server using Docker - Support for GPU acceleration for improved performance Tech stack: Dockerfile Why it matters: This Dockerfile simplifies the deployment of a DeepSpeech server, making it accessible and easy to use, especially for developers and researchers looking to integrate speech recognition into their projects.

Overview: J.A.R.V.I.S. é um assistente de voz para Linux que ouve você, pensa com o Claude Code, responde falando, e controla o computador, abrindo apps, mexendo no volume e na música, e rolando comandos. Key features: - Speech recognition and synthesis - Ability to control the computer - Customizable with Claude Code - Compatible with Linux and Python 3 - Available in Portuguese Tech stack: Python Why it matters: J.A.R.V.I.S. simplifies your interaction with technology by providing a voice assistant that can perform tasks, control your computer, and even answer your queries, all while speaking your language. Repository: NicholassweetDev/jarvis GitHub stars: 30 GitHub description: Assistente de voz para Linux que ouve você (Whisper local), pensa com o Claude Code e responde falando enquanto abre apps, mexe no volume e na música, abre sites e roda comandos.

Overview: FutureOmni is a benchmark designed to evaluate future-event forecasting from audio-visual context in multimodal language models. It focuses on assessing models' ability to predict future states based on causal logic and temporal reasoning. Key features: - **First Omni-Modal Forecasting Benchmark**: Dedicated to evaluating omni-modal future forecasting. - **Scalable Construction Pipeline**: Built using a robust LLM-assisted, human-in-the-loop pipeline. - **Comprehensive & Original**: Contains 919 videos and 1,034 multiple-choice QA pairs across 8 primary domains, all collected to ensure zero contamination. - **Challenging Evaluation**: Evaluations on 13 omni-modal and 7 video-only models reveal significant shortcomings in current systems. - **OFF Training Strategy**: Proposes the Omni-Modal Future Forecasting (OFF) strategy and provides a 7K-sample instruction-tuning dataset. Why it matters: Future forecasting from audio-visual cues is crucial for applications like personalized recommendations, educational content creation, and more, as it requires both listening and seeing capabilities. Repository: OpenMOSS/FutureOmni Primary language: Python GitHub stars: 30 GitHub description: A benchmark for evaluating future-event forecasting from audio and video context in multimodal language models

### Overview: This project is a command-line utility named **vm** that converts voice memos recorded from the microphone into text notes that can be saved in Joplin, an open-source note-taking application. ### Key features: - Converts voice memos to text notes. - Transcribes from an audio file. - Saves text to the clipboard or creates a new note in Joplin. - Supports automatic transcription for Joplin To-Dos. ### Tech stack: Shell command-line utility. ### Why it matters: This project was a proof of concept for a command-line utility that converts voice memos to text notes, which can be easily integrated into Joplin for better organization and accessibility.

Overview: This project provides a simple API for generating AI videos from text prompts, suitable for creating engaging short-form content for various social media platforms. Key features: - Text-to-Video functionality - Customizable voice options - Multiple themes available - Support for multiple languages - Duration control - Webhook delivery for video metadata Tech stack: Python 3.7 Why it matters: This tool simplifies the process of creating AI-generated videos, making it easier for developers to create engaging short-form content quickly and efficiently. Repository: SamurAIGPT/Text-To-Video-API GitHub stars: 30 GitHub description: Text to Video API generation documentation README excerpt: This repository provides a simple API for generating AI videos from text prompts, suitable for creating engaging short-form content for various social media platforms.

### Overview: LinearAdapterTrainer is an AI/ML Python library designed to fine-tune retrieval embeddings without retraining the embedding model. It uses a linear query adapter to align retrieval embeddings with user queries, providing improved retrieval performance with minimal reindexing of the corpus. ### Key Features: - **Fine-tuning without retraining**: The adapter learns a small linear transform on the query side, making it compatible with existing knowledge bases. - **Composable modules**: The library consists of two modules: a dataset generator for creating triplets and a linear adapter trainer for training the adapter. - **Triplet loss**: The adapter is trained with triplet loss, which is effective for improving retrieval performance. - **Precision@k, Recall@k, MRR, nDCG**: The adapter reports retrieval metrics such as precision, recall, MRR, and nDCG to evaluate its performance. ### Tech Stack: - **Python 3.12+**: The library is built using Python 3.12 and supports Python 3.12+. ### Why It Matters: - **Cost-effective**: The linear adapter approach is more cost-effective compared to fine-tuning the embedding model, making it suitable for scenarios where retraining is not feasible. - **Reversible**: The adapter is reversible, allowing for easy reindexing of the corpus without retraining. - **Performance Improvement**: The adapter improves retrieval performance by nudging the query embeddings closer to relevant chunks and away from irrelevant ones. ### Repository: - **Repository**: SantanderAI/linear-adapter-trainer

### Overview: DR Pixel is a page-native deep research agent designed to efficiently retrieve and analyze scientific papers, preserving the visual content and structure of pages alongside textual information. ### Key Features: - **Page-Native Retrieval**: DR Pixel renders each page of a scientific paper, integrating text and visual content. - **Vision-Language Embedding Model**: Uses Qwen3-VL-Embedding-8B to embed page images for better retrieval. - **Counterfactual Credit Reweighting**: Enhances model performance through SFT recipe training. - **Evaluation Harness**: Provides 500 questions over 1.06M pages from 99,055 scientific papers, graded by any OpenAI-compatible model. ### Tech Stack: Python ### Why It Matters: DR Pixel offers a more comprehensive approach to deep research by preserving the visual content of scientific papers, making it particularly useful for researchers and educators who need to quickly understand the full context of their documents. Repository: StarTrail-org/DRPixel GitHub Stars: 17

### Overview: This repository is a community-driven collection of creative use cases for LLM image generation capabilities, showcasing how various types of images can be generated using GPT-4o. It aims to inspire people to explore the power of multimodal AI through examples and contributions. ### Key features: - **Image Processing:** Transform black and white photos into color photos. - **Style Transfer:** Convert images into Chibi or Ghibli style. - **Logo Design:** Convert 2D logos into 3D or neon effects. ### Tech stack: Unknown ### Why it matters: This repository serves as an inspiration hub and contribution platform for exploring the potential of multimodal AI through image generation. It encourages collaboration and sharing of creative projects, aiming to make the community more collaborative and less competitive.

Overview: Aqueduct is a simple yet fully-featured AI gateway built on top of the LiteLLM Router SDK, designed for self-hosting with MCP support. Key features: - MCP server support - No SSO tax - No observability tax - No self-hosting tax - No org management tax Tech stack: Python Why it matters: Aqueduct simplifies AI gateway self-hosting by avoiding re-implementing OpenAI-compatible APIs and features, using Django for a clean and efficient implementation, and leveraging the LiteLLM Router SDK for enhanced control and routing capabilities. Repository: TU-Wien-dataLAB/aqueduct GitHub stars: 30

Overview: Vosk4Unity is a Unity module that simplifies integrating speech recognition into your applications, powered by the Vosk engine. Key features: - Easy integration of speech recognition - In-editor GUI for model management - Supports Vosk-compatible models - Quick setup process Tech stack: C# Why it matters: Vosk4Unity streamlines the process of adding speech recognition to Unity projects, making it accessible for developers without deep knowledge of speech recognition technology. Repository: Yeti47/Vosk4Unity GitHub stars: 30

Overview: DataProcessingFramework is a comprehensive framework designed for processing and filtering multimodal datasets. Key features: - Supports reading and filtering datasets. - Implements various filters and transformations for image and video data. - Provides pipelines for data transformation and validation. - Offers support for different file systems including local and S3-compatible storage. Tech stack: Python Why it matters: This framework simplifies the process of working with large datasets, enabling efficient filtering and transformation of data, making it easier to manage and analyze complex multimodal datasets.

### Overview: A Python toolkit designed to compress the KV cache in large language models, optimizing the memory bottleneck during inference. The library offers 12 different methods for compressing the KV cache, providing a unified API for optimal performance. ### Key Features: - **12 Different Methods**: Compress the KV cache in various ways. - **Godzilla KVarN/TriAttention**: Uses the Godzilla KVarN/TriAttention backend for improved performance. - **CUDA Weight Sharing**: Shares weights across multiple GPUs for better utilization. - **Multi-GPU Planning**: Plans and optimizes GPU usage for maximum efficiency. ### Tech Stack: - Python ### Why It Matters: - Optimizes the memory usage of large language models, reducing the dependency on memory-intensive KV cache. - Provides a unified API for easy integration and management of different compression methods. ### Repository: - **GitHub Repository**: aivrar/multi-turboquant - **GitHub Stars**: 28 - **GitHub Description**: Unified KV-cache compression toolkit for LLM inference: 12 Python-native methods, guarded add-on composition and routing, analytical capacity simulation, Godzilla KVarN/TriAttention with SM86/SM89 qualification, CUDA weight sharing, and multi-GPU planning. ### Installation: - Clone the repository: `git clone https://github.com/aivrar/multi-turboquant` - Install the package: `pip install -e .` - Run the dashboard: `python run_ui.py` ### Methods: | Method | Family | Transform | Bits | Compression |

Overview: 阿亮开源的 Skills 合集：短剧成片、电商详情图、儿童故事、有声绘本、配音、多人播客。一句话触发，AI 走完整 AIGC 流程。 Key features: - 一句话灵感生成的短剧成片 - 产品图去背景 + 生成多角度/多场景电商详情图 - 按年龄/时长生成儿童故事，并配音 - 故事主题生成的有声绘本动画 - 多人对话脚本 → 分角色配音 → 拼成完整播客 - 用百炼 CosyVoice v3.5 合成语音 Tech stack: Python Why it matters: 该合集为任何支持 Skill 的 AI 助手提供了一套完整的 AIGC 流程，简化了从灵感到成品的整个过程，提高了工作效率和创意产出。

### Overview: Z-Image Turbo API 한국어 가이드: 모델 ID, 장당 $0.01 요금, 파라미터, cURL·Python 예제와 OpenAI 호환 엔드포인트 설명. ### Key features: - **모델 ID**: `z-image-turbo` - **장당 가격**: $0.01 - **파라미터**: `prompt`, `n`, `size`, `model` - **cURL 예제**: `curl --request POST ...` - **Python 예제**: `import requests, time, os, os.environ; ...` ### Tech stack: Python ### Why it matters: - **모델 ID**: 가장 저렴한 이미지 모델로 사용 가능 - **장당 가격**: $0.01의 장당 요금제 - **구독**: 무료 한도 없음 - **과금액 포함**: 모든 호출에 비용이 반환됨 - **비동기 작업**: `task_id`를 받고 `GET /v1/tasks/{id}`를 폴링 ### Repository: apimart-1123/z-image-api-ko ### Primary language: Unknown ### GitHub stars: 30 ### GitHub description: Z-Image Turbo API 한국어 가이드: 모델 ID, 장당 $0.01 요금, 파라미터, cURL·Python 예제와 OpenAI 호환 엔드포인트 설명

### Overview: FLUX.2 API 日本語ガイド：モデル ID、1枚あたり $0.024 (1MP) の料金、パラメータ一覧、curl・Python の実行例と OpenAI 互換エンドポイントの解説。 ### Key features: - 1枚あたり $0.024 (1MP) の従量課金 - 最低 1 ドルからチャージでき、OpenAI 互換の `https://api.apimart.ai/v1` だけで完結します - モデル ID、1枚単価、実行例、パラメータ一覧、curl・Python の実行例 - 1MP は1枚、1秒は1枚、100万トークン単位の言語モデル - リクエストパラメータ、60 秒で開始のサンプル、実測料金、実際の費用の計算方法 ### Tech stack: - GitHub ### Why it matters: - 1 つのキーで全カタログ - 最低 1 ドル、従量課金 - 費用がレスポンスで返る - 非同期タスク前提の設計 - モデル ID とエンドポイント ### Repository: apimart-223/flux-2-api-ja GitHub stars: 3