Projects

### Overview: Iris-3B Metal is a Metal-optimized engine for text-to-image generation, designed for Apple Silicon GPUs. It includes a native macOS app, CLI, Python API, and local HTTP API, offering direct RGB pixel-space diffusion for enhanced performance and efficiency. ### Key Features: - Direct RGB pixel-space diffusion on Apple Silicon. - Independent MPS and Metal inference engine. - Native macOS app, CLI, Python API, and local HTTP API. - Supports text-to-image generation. - Optimized for performance and efficiency. ### Tech Stack: Python ### Why It Matters: Iris-3B Metal provides a direct and efficient text-to-image generation process, leveraging the capabilities of Apple Silicon GPUs for optimal performance and reduced latency. The engine's native macOS app and Python API offer a versatile and powerful toolset for developers and artists looking to integrate this technology into their workflows.

### Overview: Semantic Forge is a cutting-edge fine-tuning toolkit designed to transform general-purpose Large Language Models into domain-specific intelligence. Unlike conventional fine-tuning frameworks, it offers a cognitive sculpting environment that blends advanced NLP architectures with an intuitive, visual workflow, making the entire fine-tuning process accessible and understandable. ### Key Features: - **Adaptive Curriculum Engine**: Automatically adjusts training intensity based on real-time loss landscapes, reducing training time by up to 37% without sacrificing final performance. - **Architecture-Aware Optimization**: Recognizes and optimizes the structural differences between various transformer families, ensuring that attention heads are updated in a way that preserves the base model's linguistic priors while injecting domain-specific knowledge. - **Multilingual Fine-Tuning Framework**: Supports 47 languages across 12 writing systems, automatically adjusting tokenization strategies based on the language composition of the dataset. ### Tech Stack: Semantic Forge leverages cutting-edge NLP architectures and infrastructure, providing a comprehensive solution for fine-tuning models in various domains. ### Why It Matters: Semantic Forge addresses the three fundamental flaws of standard fine-tuning frameworks: 1. **The Black Box Problem**: It provides insights into why certain choices matter, unlike hidden configurations. 2. **The Vanilla Model Trap**: It accounts for architectural nuances across different transformer families, ensuring optimal performance. 3. **The Deployment Gap**: It offers a practical solution for evaluating, versioning, and deploying the resulting artifact safely, bridging the gap between research and deployment. With Semantic Forge, developers can easily adapt their models for specialized

### Overview: LawDIS is a language-window-based controllable dichotomous image segmentation framework designed for high-accuracy and personalized image segmentation tasks. It integrates macro and micro control modes, enabling users to generate initial masks based on language prompts and refine them using flexible local windows, respectively. ### Key Features: - **Framework Innovation:** LawDIS recasts the dichotomous image segmentation (DIS) task as an image-conditioned mask generation problem within a latent diffusion model, enabling seamless integration of macro and micro user controls. - **Dual Control Modes:** LawDIS employs a mode switcher to coordinate two distinct control modes: macro mode (LS) for language-controlled segmentation and micro mode (WR) for flexible refinement of user-defined regions. - **Flexible Adaptation:** LS and WR can operate independently or collaboratively, meeting high-accuracy personalized demands and serving as a general-purpose post-refinement tool for any segmentation model. - **Superior Performance:** Extensive evaluations on the DIS5K benchmark show that LawDIS consistently outperforms 11 state-of-the-art methods. ### Tech Stack: - Python ### Why It Matters: LawDIS offers a novel approach to image segmentation by integrating language and window-based control, making it highly versatile and effective for personalized segmentation tasks.

### Overview: This repository is a collection of papers and resources on Time Series Foundation Models (TSFMs) that have been trained from scratch, focusing on their development from pre-training to post-training. It serves as a comprehensive reference for researchers and practitioners interested in building, adapting, and deploying foundation models for time series analysis. ### Key Features: - 📄 **Datasets**: Discusses large-scale source domain datasets and the impact of target domains on model evaluation. - 🧪 **Pre-training Strategies**: Covers task-agnostic and task-specific training paradigms. - 🚀 **Post-training Methodologies**: Includes optimization via supervised, collaborative, and reinforcement-based fine-tuning. ### Tech Stack: Unknown ### Why it Matters: This repository provides a structured overview of TSFMs, focusing on their development from pre-training to post-training, making it a valuable resource for researchers and practitioners looking to build, adapt, and deploy foundation models for time series analysis.

Overview: This project provides a comprehensive API for accessing the GPT Image 2 model, offering various pricing options and parameters for developers to use in their projects. Key features: - Access to the GPT Image 2 model via the APIMart endpoint - Support for different pricing models: token-billed and cached_text_input - Flexible resolution options and customizable prompts - Batches and retries for async tasks Tech stack: Python Why it matters: This API simplifies the integration of GPT Image 2 into projects, providing a single, unified endpoint for all model interactions, making it easier for developers to manage and scale their API calls without the need for separate subscriptions or prepaid plans. Repository: apimartstev/gpt-image-2-official-api-cn Primary language: Python GitHub stars: 53 GitHub description: GPT Image 2 (official) API (`gpt-image-2-official`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

### Overview: EmbedFlow is a zero-downtime solution for migrating embedding models over existing vector indexes, ensuring that the existing retriever remains useful during the transition. ### Key Features: - Progressive embedding-model migration over existing vector indexes. - Supports migration analysis, persistent caching, background work, and serving through various vector stores (FAISS, Qdrant, pgvector, Pinecone, Milvus, Weaviate). - Provides a deterministic demo for quick setup without requiring paid services or model downloads. - Includes a script for source-checkout development, `scripts/run_demo.sh`. ### Tech Stack: - Python ### Why It Matters: EmbedFlow simplifies the process of upgrading embedding models by allowing the new model to serve over existing vector indexes while the target document vectors are materialized progressively. This approach leverages the existing index to preserve useful retrieval neighborhoods, making the transition smoother and more efficient.

### Overview: TartanAviation is an open, multi-modal dataset of terminal-area airspace operations, recorded with sensors installed inside the boundaries of two Pittsburgh-area airports: Allegheny County Airport (KAGC) and Pittsburgh-Butler Regional Airport (KBTP). The collection spans from 2020 to 2023, providing a comprehensive view of terminal airspace operations through vision, speech, and ADS-B trajectories. The dataset includes over 3.1 million images, 3,374 hours of audio recordings, and 661 days of aircraft trajectories, covering a wide range of weather and traffic conditions. ### Key features: - **Vision**: 3.1 million images in 550 videos, with aircraft bounding boxes tied to each aircraft's ICAO address, range, and type; synchronized ADS-B; per-video weather. - **Speech**: ~3,374 hours of ATC radio in 677 daily archives, with per-clip metadata; raw recordings also available. - **ADS-B**: 661 days of aircraft trajectories, raw and processed. - **Weather**: METAR reports for 2020 to 2022. ### Tech stack: Python ≥ 3.8 ### Why it matters: TartanAviation offers a valuable resource for researchers and developers interested in improving airport operations, airspace management, and aviation safety. The comprehensive dataset provides insights into real-world terminal airspace scenarios, enabling the development of more accurate and effective predictive models.

Overview: GAKG is a multimodal Geoscience Academic Knowledge Graph framework that fuses papers' illustrations, text, and bibliometric data. It currently consists of over 68 million triples and is the largest and most comprehensive geoscience academic knowledge graph. The graph is accessible via the SPARQL endpoint at [gakg/sparql](https://www.acekg.cn/sparql). Key features: - Fusion of papers' illustrations, text, and bibliometric data - Preservation of datasets and GAKG-oriented resources - Community detection, including GPCN, GACN, and their respective metrics - Knowledge representation learning, with benchmarks GA16K and WN18 - Paper knowledge extraction pipeline, utilizing BERT-based QA Tech stack: Python Why it matters: GAKG serves as a comprehensive resource for data mining and knowledge discovery in geoscience, providing valuable datasets and tools for academic research and community collaboration.

### Overview: KubePilot is a next-generation orchestration companion for Kubernetes that turns the platform into a self-aware stage for AI workloads. It translates Kubernetes' nouns into verbs, enabling platform teams and ML engineers to think in outcomes rather than pods. ### Key Features: - **Intent-Driven Mission Compiler**: Defines success criteria, KubePilot generates resource graph. - **Warm Pool Recycling (Anchors)**: Keeps GPUs, images, and weights ready, reducing cold-start latency. - **Predictive Scheduler**: Anticipates resource contention based on historical Beacon data. - **Multilingual Operator Console**: Supports multiple languages for better accessibility. ### Tech Stack: HTML ### Why It Matters: KubePilot simplifies the orchestration of AI workloads by translating Kubernetes' nouns into verbs, making it easier for teams to think in outcomes rather than pods. This approach reduces the need for manual intervention and improves the overall efficiency and reliability of AI workloads running on Kubernetes. ### GitHub Stars: 53 ### Repository: [fh6312/kubeflow-pipeline-forge](https://github.com/fh6312/kubeflow-pipeline-forge) ### Primary Language: HTML ### Disclaimer: This project is in its early stages and may not be production-ready. Use at your own risk. ### License: MIT License

Overview: This Kubernetes operator simplifies the deployment and management of Qdrant vector databases, enabling single-node and multi-node clusters with advanced features like authentication, TLS encryption, and customizable settings. Key features: - Single-node and multi-node Qdrant cluster creation and scaling - Support for API keys with read-write and read-only permissions - TLS encryption for client and peer-to-peer communication, with auto-generated and user-provided certificates - Custom Qdrant parameter support - Scheduling options management for Qdrant collections - Leader elections for high availability - Instant and scheduled snapshots storage in S3-compatible storage Tech stack: JavaScript Why it matters: This operator streamlines the deployment process for Qdrant, making it easier to manage and scale vector databases in Kubernetes environments, ensuring high availability and security through advanced features like TLS encryption and customizable settings.

Overview: A collection of basic Python modules for spoken natural language processing. Key features: - Spoken language recognition - Text-to-speech conversion - Named entity recognition - Sentiment analysis Tech stack: Python Why it matters: This project offers essential tools for developers working with spoken natural language processing, making it a valuable resource for applications requiring speech-to-text functionality and sentiment analysis.

Overview: 本项目使用 Gemini AI 转录音频视频为 SRT 字幕，通过智能切片、逐片转录和精准组装，提供时间轴精确的字幕文件。项目旨在解决直接将音视频文件发送给 Gemini AI 时时间轴不精确的问题。 Key features: - 使用 Gemini AI 转录音频视频为 SRT 字幕 - 智能切片将音视频文件分割成小片段 - 逐片转录每个片段 - 精准组装转录结果，确保时间轴的准确性 Tech stack: Python Why it matters: 该项目提供了一种高效、准确的方法来将音视频文件转录为 SRT 字幕，解决了时间轴不精确的问题，适用于需要精确时间轴的场景。

### Overview: Voxa is a macOS app that transforms the Apple TV Remote (Siri Remote) into a versatile Mac remote control. It supports voice control, allowing users to control their Macs with just a few simple gestures and commands. ### Key Features: - **Voice Control**: Control your Mac by voice through the remote's microphone. - **Button Mapping**: Map any Siri Remote button to a key, shortcut, or common action. - **App Scenes**: Different mappings for each app, automatically switching when the app comes to the front. - **Clickpad as a Trackpad**: Move the pointer and click from the couch. - **Spin the Ring to Switch Apps**: Switch between apps with a single ring notch. - **Shortcuts**: Browse and run macOS Shortcuts from the remote. - **Hold Mode**: Hold the Siri button and the mapped key stays held, working with hold-to-talk dictation tools. - **Launch at Login**: The remote works as soon as your Mac starts, quietly from the menu bar. ### Tech Stack: - **GitHub Repository**: jinzongma/apple-tv-remote-for-mac - **GitHub Stars**: 55 ### Why It Matters: - Voxa simplifies the process of controlling your Mac with the Apple TV Remote, making it more accessible and efficient. The voice control feature adds a new level of convenience, allowing users to interact with their devices without needing to physically touch them.

### Overview: PhyX is a large-scale benchmark designed to evaluate models' physical reasoning capabilities. It includes 3,000 multimodal questions covering 6 reasoning types across 25 sub-domains and 6 core domains, including thermodynamics, electromagnetism, mechanics, modern physics, optics, and wave acoustics. PhyX challenges models to integrate visual cues with physical laws, making it distinct from general-purpose benchmarks. ### Key Features: - **Dataset**: PhyX comprises 3,000 multimodal questions. - **Reasoning Types**: 6 types of reasoning across 25 sub-domains and 6 core domains. - **Core Domains**: Thermodynamics, electromagnetism, mechanics, modern physics, optics, and wave acoustics. - **Visual Scenarios**: Realistic, high-fidelity visual scenarios presented through multimodal inputs. ### Tech Stack: - Python ### Why It Matters: PhyX is crucial for evaluating models' physical reasoning abilities, making it essential for researchers and developers in the field of artificial intelligence and machine learning.

Overview: This project aims to build a RAG-based chatbot that utilizes OceanBase or OceanBase SeekDB's AI capabilities, specifically focusing on hybrid search and AI embedding, to enhance the accuracy of its answers. Key features: - Text Embedding Service - Database for Document Vectors and Structured Data - Large Language Model Agents - Streamlit Chat Interface Tech stack: Python Why it matters: This project showcases the integration of OceanBase's AI capabilities with a chatbot, leveraging vector retrieval to provide more accurate answers to users' questions. Repository: ob-labs/ChatBot Primary language: Python GitHub stars: 55 GitHub description: ChatBot, show how to implement a RAG based on OceanBase or OceanBase seekdb AI capabilities especially hybrid search and AI embedding.

### Overview: This repository is the official implementation of the paper "Physics Informed Distillation for Diffusion Models," accepted by Transactions on Machine Learning Research (TMLR). It introduces Physics Informed Distillation (PID), which leverages a student model to approximate the solution of the ODE system corresponding to a teacher diffusion model. PID achieves comparable performance to recent distillation methods while eliminating the need for synthetic dataset generation during the distillation process. ### Key Features: - Physics Informed Distillation (PID) - Uses a student model to approximate the solution of the ODE system - Eliminates the need for synthetic dataset generation during distillation - Comparable performance to recent distillation methods ### Tech Stack: Python ### Why It Matters: PID simplifies the distillation process for diffusion models, making it more accessible and efficient. It eliminates the need for extensive dataset generation, making it easier to use and deploy. ### GitHub Stars: 55 ### Repository: pantheon5100/pid_diffusion

Overview: agenticow is a Git for Agent Memory that branches vector memory, providing 83x faster and 3000x smaller snapshots compared to traditional full-copy methods. It is optimized for embedded multi-agent memory and is designed to be used with JavaScript. Key features: - Branching memory in ~0.5 ms / 162 bytes, independent of base size. - Exact read-through queries with parent ∪ edits, child wins on id collisions and honors deletes. - Built for embedded multi-agent memory. Tech stack: JavaScript Why it matters: agenticow provides a more efficient and scalable way to manage memory for agents, reducing the cost of branching and checkpointing, and enabling parallel experimentation and personalization. Repository: ruvnet/agenticow GitHub stars: 55 GitHub description: Git for Agent Memory: Copy-On-Write vector branching for embedded multi-agent memory (83x faster, 3000x smaller snapshots)

### Overview: BabyTalk is a fully offline, no-cloud, speech-to-text and text-to-speech system for ESP32 microcontrollers. It optimizes performance with high-performance 4/8-bit int kernels and a finetuned STT model for noisy environments. Features include a data generation/field capture/model finetuning framework, and AtomVM support for C, MicroPython, and Erlang/Elixir. ### Key Features: - **Speech to Text (4-bit model)**: 6.0 MB of flash, 84.5 KB of internal RAM, 0.2 MB of flash for transcribing 4 seconds of audio. - **Text to Speech (nano voice)**: 1.9 MB of flash, 0.5 MB of internal RAM, 1.5 MB of flash for pronunciations. - **Performance and Memory**: Speech to text requires 6 MB of flash and 84.5 KB of internal RAM; text to speech adds 1.9 MB of flash. - **Models**: Supports 8-bit and 4-bit models, with the 8-bit model taking up 9.8 MB of flash and the 4-bit model taking up 6.0 MB of flash. - **PSRAM Usage**: The 8-bit model uses 0.2 MB of flash, while the 4-bit model uses 0.5 MB of flash. The text to speech model uses 1.5 MB of flash. - **Wake Phrase**: Supports any phrase as a wake word, with the

Overview: 3DCF / doc2dataset is a token-efficient document layer designed for LLMs, featuring NumGuard numeric integrity and multi-framework exports for RAG and fine-tuning. Key features: - Document layer standard with deterministic macro-cells - NumGuard numeric integrity for per-cell number hashes - Token-efficient contexts with reduced token size - Reusable QA/Summary/RAG samples and multi-framework exports - Publishable Rust core and Rust-native tools Tech stack: Rust Why it matters: This open-source project provides a reproducible document layer that feeds RAG and fine-tuning pipelines, ensuring numeric integrity and token efficiency, making it suitable for ML/AI platform teams, fintech/ regulatory / analytics teams, and LLM researchers / OSS devs. Repository: 3DCF-Labs/doc2dataset GitHub stars: 54

### Overview: ModelGate is a contract-aware AI control plane that routes requests to the optimal model based on customer contracts and SLA/privacy/routing constraints. ### Key Features: - **Contract Compliance**: Ensures compliance with customer contracts by routing requests to the appropriate model. - **Cost Optimization**: Automatically routes requests to the cheapest model for simple queries and the most expensive model for complex ones. - **Model Selection**: Routes requests to the optimal model based on the customer's contract and SLA requirements. ### Tech Stack: TypeScript ### Why It Matters: Over 30 new LLMs launched in the past month alone. No team has time to evaluate them all - so they pick one premium model and send everything to it. The result: 50-90% of enterprise AI spend is wasted on over-provisioned models, and premium models consume 180x more energy per query than small ones. ModelGate fixes this. You change one line of code - your `base_url` - and we handle model selection, contract compliance, and cost optimization automatically. This results in significant cost savings and better resource utilization. ### Repository: Aaryan-Kapoor/ModelGate-Hackathon ### GitHub Stars: 54 ### Primary Language: TypeScript ### GitHub Description: 🏆Winning Project | ModelGate is a contract-aware AI control plane that ingests customer contracts, extracts SLA/privacy/routing constraints, and generates an OpenAI-compatible endpoint that automatically routes every request to the optimal model. Simple queries go to cheap models