Projects

Overview: The FLUX.2 Pro API (flux-2 / flux2) offers flat per-unit pricing for image generation, with rates starting at $0.024 for 1MP, $0.036 for 2MP, and $0.048 for 3MP. This API is compatible with OpenAI's APIMart gateway, allowing for cost-effective image creation. The minimum top-up is $1, and the pricing is subject to change. Key features: - **Pricing**: $0.024 for 1MP, $0.036 for 2MP, $0.048 for 3MP. - **Cost at Scale**: $2.4 for 100 units, $24 for 1,000 units. - **API Gateway**: Compatible with OpenAI's APIMart gateway, providing a cost-effective alternative to traditional APIs. Tech stack: Shell Why it matters: The FLUX.2 Pro API is an efficient and cost-effective solution for generating high-quality images, making it ideal for developers and content creators looking to streamline their workflows without the high costs associated with traditional APIs.

### Overview: LIDIA is a lightweight image denoising network that adapts to input images, achieving near state-of-the-art results with minimal parameters. ### Key features: - **Adaptive Denoising**: LIDIA automatically adjusts to the characteristics of the input image. - **Multi-Scale Processing**: Combines processing of patches from different scales for better results. - **Lightweight Implementation**: Uses a small number of parameters, making it highly efficient. ### Tech stack: Python ### Why it matters: LIDIA stands out for its ability to achieve high-quality denoising with minimal computational resources, making it highly effective for real-time applications and resource-constrained environments.

Overview: LLM Inspector is a tool designed to monitor and optimize LLM inference processes, providing detailed insights into GPU memory usage and model configurations. Key features: - Measured quantization savings - Projected optimization opportunities - Detailed system observability and model internals Tech stack: Python Why it matters: LLM Inspector offers a unique approach by providing measured data alongside optimization projections, allowing developers to understand and optimize their LLM inference processes effectively. Repository: helasaoudi/llm-inspector GitHub stars: 76

### Overview: An opinionated atlas of model post-training across modalities: RL, rewards, alignment, distillation, adaptation, agents, and their connections. ### Key Features: - **7 Tracks**: R1 Full-Chain, GRPO Family, Reward & Verifiers, MoE RL Stability, Paradigm Frontier, Unified Theory, Post-Training Stability. - **Cross-Linked Papers**: Every work is read, classified, and cross-linked; most are connected back to a Chinese deep-dive by [hscspring](https://yam.gift). - **Chronological Timeline**: Ideas are kept in publishing order rather than reorganized after the fact. ### Tech Stack: Python ### Why It Matters: This repository maps the modern model post-training landscape, focusing on RL, reward, SFT × RL, distillation, and distribution alignment. It provides a curated view of the evolving mechanisms used in post-training tasks, emphasizing their practical applications and connections between different modalities.

### Overview: This project aims to develop a multilingual speech recognition model for Indonesian, Javanese, and Sundanese languages, aiming to improve the performance across multiple languages without sacrificing model performance. ### Key features: - Multilingual Speech Recognition Model - Indonesian, Javanese, and Sundanese Datasets - Wav2vec 2.0 Framework for Speech Recognition - Comparison with Google Speech To Text - Performance Evaluation Metrics: Word Error Rate (WER) ### Tech stack: Python ### Why it matters: The project addresses the need for multilingual speech recognition in Indonesia, where over 700 languages are spoken. By developing a multilingual model, it can significantly improve the performance across multiple languages without the need for separate models, thus making it more practical and efficient.

Overview: This project is designed to help developers prepare for the AWS Certified AI Practitioner (AIF-C01) exam. Key features: - Comprehensive exam preparation materials - Detailed explanations for each exam question - Interactive practice tests with immediate feedback Tech stack: Python Why it matters: This project is essential for anyone aiming to pass the AWS Certified AI Practitioner (AIF-C01) exam, providing a thorough understanding of the necessary skills and knowledge required.

Overview: This repository contains the official PyTorch implementation of DAVI, a method for solving noisy inverse problems using diffusion prior-based amortized variational inference. The framework allows efficient posterior sampling with a single evaluation of a neural network and enables generalization to unseen measurements without the need for test-time optimization. Key features: - Efficient posterior sampling with a single neural network evaluation - Generalization to both seen and unseen measurements - Five image restoration tasks: Gaussian deblur, 4x Super-resolution, Box inpainting, Denoising, and Colorization - Two benchmark datasets: FFHQ and ImageNet Tech stack: The repository uses Python, PyTorch, and other libraries such as Accelerate, EMA PyTorch, Matplotlib, Piq, Scikit-Image, PyTorch-FID, and Wandb. Why it matters: DAVI is a significant contribution to the field of image restoration and inverse problems, offering a practical solution for efficiently solving noisy inverse problems with minimal computational overhead.

Overview: A Toolkit for Distributional Control of Generative Models Key features: - Supports multiple generative models - Implements distributional control techniques - Provides easy-to-use API for developers Tech stack: Python Why it matters: This toolkit offers a comprehensive solution for developers working with generative models, enabling them to easily implement distributional control techniques, thereby enhancing the performance and accuracy of their models.

### Overview: Nano Banana Pro is a powerful image generation and editing API offering a cost-effective alternative to Google's Gemini 3 Pro. With a default price of $0.03 per image (4K $0.04), it provides a single key access to over 300 models, including Nano Banana Pro, without requiring a subscription or upfront payment. The API supports OpenAI-compatible endpoints, allowing for seamless integration with various applications. ### Key Features: - **One Key, Entire Catalog**: Use the same `https://api.apimart.ai/v1` base URL and `Authorization` header to reach Nano Banana Pro and over 300 other models. - **$1 Minimum, Pay as You Go**: No subscription required; pay for each call, and the charge is reflected in the response. - **Charge Back to the Response**: Every completed task reports the amount billed, and a failed task is refunded. - **Async by Design**: Submit tasks, and poll the `GET /v1/tasks/{id}` endpoint for updates. ### Tech Stack: The API is built on Python, utilizing the `requests` library for HTTP requests and handling the API endpoints. ### Why It Matters: - **Cost-Effective**: Offers a cost-effective alternative to Google's Gemini 3 Pro, with a default price of $0.03 per image (4K $0.04). - **OpenAPI Compatibile**: Provides an OpenAI-compatible endpoint, making it easy to integrate with applications that support OpenAI-compatible APIs. - **No Subscription

Overview: This repository contains curated AI prompts designed for developers using AI coding agents. Unlike typical lists, these prompts are built to verify work and stop AI from hallucinating solutions. They are built with senior engineer workflows and include evidence, tooling, and verification to ensure accurate results. Key features: - Disciplined workflows for developers using AI coding agents. - Senior-engineer-level prompts that verify work and stop AI from hallucinating. - Tool-agnostic, works with Claude, ChatGPT, Copilot, Cursor, and any other agent that reads prompts. - Verification-first approach, ensuring agents are honest. - Two formats: one-pagers for quick tasks and multi-section prompts with hard constraints for big, risky work. - Self-contained, everything needed in one copy-paste block. - Auto-appends contributors to the CONTRIBUTORS.md file. Tech stack: HTML Why it matters: These prompts provide disciplined workflows that ensure accurate and verified results, making AI coding agents more reliable and trustworthy for developers. Repository: shauryagangrade/awesome-ai-prompts GitHub stars: 76

### Overview: GeminiWithFiles is a powerful library designed to integrate the Gemini API into Google Apps Script, making it easier to process files and unstructured data. It simplifies the implementation of AI-driven solutions, allowing developers to create autonomous agents using Agent Skills. ### Key Features: - Seamless integration with the Gemini API - Support for uploading files, generating structured content, and executing code - Detailed token usage tracking - Support for Agent Skills, enabling dynamic, autonomous workflows ### Tech Stack: JavaScript ### Why it Matters: By streamlining interactions with the Gemini API, developers can quickly build AI-driven solutions across various domains without the need for extensive coding, making the process more efficient and accessible.

Overview: ReportAI is a Python-based tool designed to generate comprehensive PDF research reports using AI models like GPT, Claude, and Llama. It ensures detailed reports with proper sections, tables, and figures, reducing repetition and making the content clear and concise. Key features: - Generates detailed research reports from AI-generated conversations. - Produces sections with headers, tables, and figures. - Customizable via environment variables. - Easy to use with minimal setup. Tech stack: Python Why it matters: ReportAI simplifies the process of generating professional research reports, making it easier for developers to create detailed and informative documents without the need for extensive coding skills. Repository: AdirthaBorgohain/reportAI Primary language: Python GitHub stars: 75 GitHub description: Generate full fledged PDF reports using LLMs like GPT, Claude, Llama README excerpt:

Overview: CodeFox is an AI full-stack project generator that creates dynamic, interactive pages directly on your local machine. Key features: - It asks before it builds, providing audience and visual direction. - The preview is the file, allowing you to see the result immediately. - Every turn is a commit, keeping track of your work history. - The project remembers previous turns, allowing you to restyle without rewriting. - It ships it, providing options to publish to Vercel, download a zip, or export a PDF. Tech stack: TypeScript Why it matters: CodeFox simplifies the process of creating dynamic, interactive web pages, making it accessible to developers without requiring a deep understanding of the underlying technologies. Repository: CodeFox-Repo/codefox GitHub stars: 75 GitHub description: Where the fox came from — the original AI full-stack project generator with multi-agent codegen and live preview.

Overview: Dewy is an opinionated knowledge extraction and semantic retrieval tool designed specifically for Gen AI applications. It simplifies the process of extracting knowledge from documents and implementing semantic search capabilities. Key features: - Extracts knowledge from documents - Implements semantic search over extracted content - Supports opinionated configuration for better application integration Tech stack: Python Why it matters: Dewy streamlines the process of building AI agents and RAG applications, focusing on the most critical aspects of knowledge extraction and semantic search to help developers focus on their core tasks. Repository: DewyKB/dewy Primary language: Python GitHub stars: 75 GitHub description: Opinionated knowledge extraction and semantic retrieval for Gen AI applications. README excerpt: <!-- Improved compatibility of back to top link: See: https://github.com/othneildrew/Best-README-Template/pull/73 --> <br /> <div align="center"> <a href="https://github.com/DewyKB/dewy"> <img src="https://github.com/DewyKB/dewy/raw/main/images/logo.png" alt="Logo" width="160" height="160"> </a> <h3 align="center">Dewy - The Knowledgebase for AI</h3> <p align="center"> Opinionated knowledge extraction and semantic retrieval for Gen AI applications. <br /> <a href="https://github.com/DewyKB/dewy"><strong>Explore the docs »</strong></a> <

### Overview: This project profile details the DeepSeek V4.1 Flash model optimized for two NVIDIA DGX Spark computers. It achieves a significant 2x speedup in decoding and prefilling, with a 2.1x increase in decode speed and a 2.3x boost in prefill speed. The model is designed to be ready for AI agents, with a 2.4x faster prefill speed and a 2.2x faster needle-in-a-haystack performance. The project includes a display-backed KV experiment and is optimized for both decode and prefill speeds. ### Key Features: - **2x Speedup in Decoding and Prefilling**: The model runs 2x faster in decoding and prefilling tasks compared to the standard vLLM setup. - **2.1x Faster Decode Speed**: The decode speed increases from 27 tok/s to 57 tok/s. - **2.3x Faster Prefill Speed**: Prefill speeds increase from 925 tok/s to 2,179 tok/s. - **2.2x Faster Needle-in-a-Haystack Performance**: The needle test with a 64K-token prompt runs 2.2x faster. - **4 Concurrent Requests**: Up to 4 requests can be processed simultaneously, reaching an aggregate throughput of 140 tok/s on code. - **Prefix Cache**: A second question about the same document has a TTFT of about 1 s, and cached prompts go to disk, reducing the TTFT to about

Overview: A powerful UI for creating images using Stable Diffusion and fine-tuned models. Key features: - Supports Stable Diffusion and models like SDXL, Stable Diffusion 1.5, DreamShaper, Kandinsky-2, DeepFloyd IF, OpenJourney, Analog Diffusion, PortraitPlus, Waifu Diffusion, HASDX, Deliberate, Vintedois Diffusion, WavyFusion, Eimis Anime Diffusion, RPG, Realistic Vision v2.0, Opengen, Cyberrealistic, Meinamix, Edge of Realism, Lyriel, BRA, Babes, Neverending Dream, AbsoluteBeauty, Realistic Vision v3.0, Illuminutty Diffusion, Realistic Vision v5.1. - Upscaling / facefixing from Codeformer, Real-ESRGAN, Swin2SR, GFPGAN, Swinir, Rundalle-sr, Esrgan. - Enhances prompts using AI via DistilGPT2 Stable Diffusion V2 model. - Supports Image2Image on a subset of models. - Enhances a FMX TImage directly from the client into the Delphi IDE. - Built with Embarcadero Delphi using the FireMonkey framework, works on Windows, macOS, and Linux with a single codebase and single UI. - Supports REST integration with Replicate.com for Generate Image functionality. - Requires Skia4Delphi to compile. Tech stack: Embarcadero Delphi, FireMonkey. Why it matters

### Overview: GLM-5.2, a 744 billion parameter mixture of experts model, is a pure C inference engine that has been quantized to int4, streamed from disk, and deployed on a machine with 16 GB of RAM. This project showcases a compact and efficient model that can run on a laptop without the need for a GPU, making it accessible for developers and researchers. ### Key Features: - **Quantization to int4**: The model is converted to 4-bit precision, reducing memory usage significantly. - **Expert streaming**: Experts are loaded from disk on demand, keeping the working set size minimal. - **Pinned RAM and LRU cache**: Experts are placed in pinned RAM and an LRU cache to optimize memory usage. - **GPU-free inference**: The model can run on a machine with no GPU, making it ideal for resource-constrained environments. - **Compressed KV cache**: Multi-head latent attention is combined with a compressed KV cache for efficient computation. - **Sparse attention indexer**: A sparse attention indexer reduces the number of operations required for attention calculations. - **Tiering system**: Experts are placed across different tiers (pinned RAM, LRU cache, VRAM, disk) to manage memory usage effectively. - **OpenAI compatible server**: The model is served by an OpenAI-compatible server with a proper scheduler and backpressure mechanism. - **Verification and correctness checks**: The model is validated against a PyTorch reference implementation, ensuring accuracy and correctness. ### Tech Stack: Pure C ### Why It Matters: This project

### Overview: This repository contains two sample projects built for the FourthBrain Generative AI Workshop. These projects showcase how prompt engineering and fine-tuning can be used to optimize the output of large pre-trained models, fine-tune a Large Language Model (LLM) or text-to-image diffusion model using a data-centric approach, and deploy a Generative AI application to Hugging Face. ### Key Features: - AiconOS: A Stable Diffusion project that generates accurate and vibrant iOS icons based on textual descriptions. - GenerAds: A fine-tuned BLOOM LLM marketing assistant that can generate catchy advertisements for products. ### Tech Stack: Jupyter Notebook ### Why it Matters: These projects provide hands-on experience in building, shipping, and sharing end-to-end Generative AI applications, demonstrating the practical application of prompt engineering, fine-tuning, and deployment in the context of the workshop.

### Overview: ChronoMind is a lock-free concurrent HNSW vector index designed for AI agents, providing wait-free reads and lock-free writes. It includes a temporal memory layer that manages vectors' importance, decay rate, and access history, ensuring efficient and scalable performance. ### Key Features: - **Lock-Free Reads**: Epoch-pinned searches are performed with atomic pointer loads, ensuring wait-free access. - **Lock-Free Writes**: Publication is a single compare-and-swap, with retries only when another writer has succeeded, guaranteeing progress. - **Temporal Memory Layer**: Enhances search by blending geometric distance with recency, suitable for agent use cases. ### Tech Stack: Rust 1.82+ ### Why It Matters: ChronoMind stands out by offering a lock-free concurrent HNSW vector index with a temporal memory layer, making it highly efficient and scalable, particularly in scenarios requiring concurrent reads and writes. It outperforms established solutions like usearch and FAISS in terms of recall and search throughput, while maintaining a low latency footprint.

### Overview Kyros is an open-source memory operating system designed specifically for AI agents. It provides three biologically-inspired memory types (episodic, semantic, and procedural) with built-in features like cryptographic integrity, Ebbinghaus decay curves, and automatic contradiction resolution. ### Key Features - **Three Memory Types**: Episodic, Semantic, Procedural - **Cryptographic Integrity**: SHA-256 hashing, Merkle tree proofs - **Ebbinghaus Decay**: Natural forgetting based on category-specific decay rates - **Belief Propagation**: Automatic conflict resolution, confidence scoring - **High Performance**: <20ms average recall latency, <50ms memory storage - **Framework Integrations**: LangChain, LlamaIndex, AutoGen, CrewAI, Vercel AI SDK - **Model Agnostic**: Compatible with any LLM provider ### Tech Stack - Python ### Why It Matters Modern AI agents are stateless and fragile. Kyros provides a complete memory system in a single SDK, ensuring secure, self-correcting, and persistent memory with cryptographic integrity and natural forgetting curves. This solution is model-agnostic and compatible with popular AI frameworks, making it a game-changer for developers looking to improve the performance and security of their AI agents.