Projects

Overview: The AI-Powered Healthcare Intelligence Network is an AI-driven system offering disease prediction, drug recommendations, heart disease risk assessment, and an AI medical chatbot. It uses ML, NLP, and LLMs to provide accurate diagnoses, insights, and recommendations, enhancing healthcare accessibility, efficiency, and decision-making. Key features: - Disease Prediction & Medical Recommendation - AI-Powered Drug Recommendation - Heart Disease Risk Assessment - Medibot - AI Health Assistant Tech stack: Jupyter Notebook Why it matters: This platform revolutionizes healthcare by leveraging advanced AI technologies to improve disease prediction, personalized medical recommendations, and AI-driven insights, ultimately enhancing patient care and healthcare efficiency.

Overview: This project focuses on deploying a machine learning pipeline using Docker and FastAPI. The trained model for classifying loan borrowers as defaulters or not is included from a repository. Key features: - Dockerized server for deploying machine learning models - FastAPI framework for building the API - Pydantic library for representing loan data points - Loading and using a trained machine learning pipeline - Endpoint for making prediction requests Tech stack: Python Why it matters: This project demonstrates the practical application of deploying machine learning models using Docker and FastAPI, making it easier to scale and maintain the model in production environments.

### Overview: GPT Image 2 Hub is a curated atlas of 3798 works across 443 themes and 692 packages, featuring 3,800+ GPT-Image-2 generations. This repository provides full copy-paste prompts for OpenAI's GPT-Image-2 model, making it a valuable resource for designers, illustrators, content creators, and researchers. ### Key features: - **Works**: 3,800+ GPT-Image-2 generations across 443 themes and 692 packages. - **Prompt Details**: Every work includes its full prompt in both English and Chinese. - **Category Navigation**: Browse by category or theme to find specific prompts. - **Live Preview**: Click any work to open the live page with the full prompt. - **Prompt Reproduction**: Copy the prompt into ChatGPT Images 2.0 to reproduce or remix the output. ### Tech stack: JavaScript ### Why it matters: GPT Image 2 Hub is essential for designers and illustrators looking for reproducible style references, prompt engineers needing multilingual text rendering, content creators seeking scalable imagery, and researchers aiming to benchmark GPT-Image-2 against specific creative constraints.

### Overview: This repository contains a cross-modal benchmark for industrial anomaly detection in the fused magnesium smelting process. The benchmark includes 3 hours of synchronized video and three-phase alternating current data from different production batches. The dataset is used to explore the internal features of each modality through self-attention and cross-attention, enabling effective anomaly detection in a typical industrial process. ### Key features: - Synchronized video and three-phase alternating current data from different production batches. - Cross-modal information exploitation for anomaly detection. - Novel FMF Transformer (FmFormer) for efficient feature extraction. - Pixel-level and class-level annotations for comprehensive dataset. - Available for download via Google Drive and Baidu Disk. ### Tech stack: Python ### Why it matters: This benchmark provides a valuable resource for researchers and practitioners in the field of industrial anomaly detection, offering a practical tool for evaluating and improving detection algorithms in industrial settings.

Overview: This repository provides an example of using IBM's Watson's Streaming Speech to Text websockets interface for real-time transcription, written in Python. Key features: - Real-time speech transcription - Uses websockets for streaming - Written in Python Tech stack: Python Why it matters: This example demonstrates how to integrate Watson's speech-to-text service for real-time transcription, which can be useful for applications requiring continuous audio processing.

Overview: ZimaOS Blue is a fully open-source, auditable, vendor-neutral, and production-ready agent runtime and toolkit designed for bold developers. It is built to run private, self-hosted agents on any x86-compatible device, including Raspberry Pi, Windows, and macOS, with a memory footprint as low as 19 MB. Key features: - **Out-of-the-Box**: Ready to use with minimal setup. - **Open-Source**: Fully open-source and auditable. - **Universal**: Vendor-neutral, supporting a wide range of devices. - **Production-Ready**: Designed for performance and reliability. - **Local-First**: Runs on local hardware, ensuring privacy and security. Tech stack: Go Why it matters: ZimaOS Blue offers a flexible and customizable solution for developers looking to create and deploy local AI agents, ensuring they can ship agents with zero friction and maintain complete control over their applications. Repository: IceWhaleTech/ZimaOS-Blue GitHub stars: 29 GitHub description: ZimaOS Blue - A Local-First Agent Runtime for Bold Builders. Out-of-the-Box, Open-Source, Universal, Vendor-Neutral

Overview: This project aims to create a simple AI assistant named Jarvis using Python and Google Calendar API to manage events and provide information. Key features: - Integrates Google Calendar API to retrieve event information. - Utilizes pyttsx3 for offline speech synthesis. - Implements conditional logic to respond to specific speech commands. Tech stack: Python Why it matters: This project demonstrates basic AI capabilities using Python and Google Calendar API, showcasing how to automate tasks and provide information through speech. Repository: Ishan7390/Jarvis_AI Primary language: Python GitHub stars: 29 GitHub description: A simple AI assistant named Jarvis using Python and Google Calendar API.

Overview: Transcribe is a command-line application designed to convert images and audio files into plain text, leveraging Google Cloud services for text extraction. Key features: - OCR (Optical Character Recognition) for extracting text from images - Speech recognition for converting audio files into text Tech stack: Python Why it matters: Transcribe simplifies the process of extracting text from various media formats, making it accessible and useful for developers and content creators alike. Repository: Jaymon/transcribe Primary language: Python GitHub stars: 29 GitHub description: Convert images or audio files to plain text on the command line

Overview: Runner is a lightweight, CPU/CUDA/Metal-compatible inference engine that serves, scores, and trains LoRA models using quantized weights, ensuring byte-reproducibility. It supports OpenAI and Anthropic APIs and can be run on various platforms. Key features: - Runs on ordinary processors, NVIDIA graphics cards, and Apple Silicon. - Speaks the language of your tools (GGUF, OpenAI, Anthropic). - Measures performance accurately, comparing favorably with published models. - Can handle models of up to 30B parameters. - Leaves memory usage low, even when running large models. - Can be signed and replayed to verify results. Tech stack: C Why it matters: Runner simplifies the process of working with large models, ensuring compatibility and accuracy across different platforms, making it a valuable tool for developers and researchers.

Overview: This project implements a Dual-Level Adaptive Incongruity-Enhanced Model (DAIE) for detecting sarcasm in multimodal data, specifically leveraging text and image information. Key features: - Token-level contrastive learning with patch-based reconstructed images - Graph-level adaptive contrast learning with negative pair similarity weights Tech stack: Python Why it matters: This model improves upon existing methods by addressing limitations in image encoder encoding and negative correlation issues in graph-level feature extraction, leading to better performance in sarcasm detection.

### Overview: This repository serves as a comprehensive listing of known open-source voice tools, organized by their position within the voice stack. It aims to simplify the understanding of these tools for organizations, businesses, and individuals looking to integrate voice technology into their technical architectures. ### Key Features: - **Transcription**: Provides a platform for hosting open speech datasets, corpora, and other resources. - **Speech to Text**: Features include Kaldi Automatic Speech Recognition toolkit and Flowtron, a Tacotron-based speech synthesis tool. - **Intent Parsing**: Tools for parsing and resolving intents in natural language. - **Text to Speech**: Offers a generative text-to-speech tool based on the WaveNet architecture. - **Chatbots and Conversational UI Tools**: Includes MindMeld, a Python-based framework for building advanced conversational AI platforms. - **Voice Assistant Wrappers**: Provides wrappers for integrating voice assistants into existing applications. ### Tech Stack: TeX ### Why it Matters: Understanding the various components and tools within the voice stack is crucial for organizations looking to leverage voice technology effectively. This repository serves as a valuable resource for developers and technical teams aiming to integrate these tools into their projects and applications.

Overview: RudraDB-Opin is a relationship-aware vector database designed for learning, tutorials, hackathons, and enterprise POCs. It offers 100,000 vectors and 50,0000 relationships, including auto-dimension detection, auto-relationship detection, and auto-optimization, making it perfect for AI development. Key features: - 100,000 vectors - 50,000 relationships - Auto-dimension detection - Auto-relationship detection - Auto-optimization search - Auto-intelligent knowledge graph construction - Multi-hop discovery - No usage tracking - Production-quality code Tech stack: Python Why it matters: RudraDB-Opin is the world's first relationship-aware vector database, offering a seamless and efficient way to learn and develop AI projects without the need for manual configuration. Its auto-detection and auto-optimization features make it ideal for both personal and enterprise use cases. Repository: Rudra-DB/rudradb-opin-examples GitHub stars: 29

### Overview: Mili is an advanced artificial assistant designed to make your life easier. Powered by the Meta Wit model and Google Bard, Mili can perform a wide range of tasks, including playing songs, setting reminders, translating queries, and more. It offers an immersive conversation experience and can navigate to desired locations, check weather forecasts, and manage Wi-Fi and Bluetooth connectivity. ### Key Features: - **Task Execution**: Mili can perform various tasks such as playing songs, setting reminders, translating queries, and more. - **Mapping System**: Enables hands-free navigation to desired locations and access information about the user's current whereabouts. - **Entertainment Features**: Offers features like singing songs, riddles, fun facts, motivational quotes, jokes, and personalized movie recommendations. - **Grocery List Management**: Allows users to create and modify grocery lists for efficient management of items. - **Security and Privacy**: Secure storage of user secrets and passwords using advanced encryption and hashing techniques. - **Settings Control**: Users can control explicit content through comprehensive settings for a personalized and safe experience. ### Tech Stack: - **Programming Languages**: Python - **GitHub Stars**: 29 ### Why It Matters: Mili is designed to be a versatile and secure assistant that enhances user productivity and enjoyment. Its advanced features make it a valuable tool for managing tasks, entertainment, and everyday needs, ensuring a seamless and enjoyable user experience.

### Overview: wavrail 是一个面向 Python 的流式语音识别工具包，专为音频处理和自动语音识别（ASR）设计。它提供了一系列模块化组件，包括音频输入输出、特征提取、CTC解码、同步发射与强制对齐、指标与文本规范化、流式管道以及可选的声学模型。 ### Key features: - **音频 I/O**: 支持 PCM16、PCM32 和 Float32 格式的 WAV 文件，严格校验头部，拒绝不支持的编码。 - **特征提取**: 分帧、加窗、预加重、功率谱、Mel 滤波组、对数 Mel、CMVN 与一阶/二阶差分。 - **CTC 解码**: 贪心解码与前缀束搜索，可控的束宽与剪枝策略。 - **同步发射与强制对齐**: 流式发射循环与基于 Viterbi 的强制对齐。 - **指标与文本规范化**: 中文 CER、英文 WER、编辑对齐与文本预处理。 - **流式管道**: 有界前瞻的管道、回压、取消语义与实时因子统计。 - **可选声学模型**: 基于 CPU-only PyTorch 的小型 GRU+CTC 模型，附带合成数据训练示例。 ### Tech stack: NumPy ### Why it matters: wavrail 是一个专为流式语音识别设计的工具包，提供了一

### Overview For many **genAI solutions** in the **clinical domain**, doctor-patient transcription is a key task. It becomes especially difficult in clinical settings when the language is **code-switched**, meaning multiple languages are mixed during conversation. This is common in multilingual environments, particularly in healthcare. To address this challenge, we fine-tuned the Whisper Tiny model to **translate** **code-switched Bengali and English speech** into one language, making it easier for tasks like analysis, record-keeping, or integrating with other AI models. After fine-tuning, we obtained the [MediBeng Whisper Tiny](https://huggingface.co/pr0mila-gh0sh/MediBeng-Whisper-Tiny) model. The dataset, [MediBeng](https://huggingface.co/datasets/pr0mila-gh0sh/MediBeng), has been created using **synthetic data generation** to simulate code-switched conversations, allowing for more robust training of the model. This dataset was used in the fine-tuning process. This solution is designed to transcribe and **translate** **code-switched Bengali-English** conversations **into English** in clinical settings, helping practitioners process the information more effectively and use it for patient records or decision-making. We achieved great results after fine-tuning the model. ### Key Features - **Base model**: openai/whisper-tiny - **Fine-tuned for**: Translation task (code-mixed Bengali-English → English) - **Domain**: Clinical/M

Overview: This project focuses on interpreting CLIP with Hierarchical Sparse Autoencoders (MSAE) at the International Conference on Machine Learning 2025, aiming to improve the interpretability and control of vision-language models like CLIP. Key features: - MSAE: A new architecture that optimizes both reconstruction quality and sparsity simultaneously, enabling direct optimization of both metrics without compromise. - CLIP: A large-scale vision-language model fundamental in modern systems. - Cosine Similarity: A metric to evaluate the similarity between representations. - Fraction of Variance Unexplained: A metric to evaluate the explainability of the model. - Semantic Concepts: Over 120 semantic concepts extracted from CLIP's representation for downstream tasks. Tech stack: Jupyter Notebook Why it matters: MSAE establishes a new state-of-the-art Pareto frontier between reconstruction quality and sparsity for CLIP, achieving high cosine similarity and low fraction of variance unexplained while maintaining high sparsity. This makes MSAE a valuable tool for interpreting and controlling CLIP in downstream tasks like CelebA. Repository: WolodjaZ/MSAE Primary language: Jupyter Notebook GitHub stars: 29 GitHub description: Interpreting CLIP with Hierarchical Sparse Autoencoders (ICML 2025)

### Overview Barq-DB v2 is a Rust-based retrieval system designed for modern AI workloads. It combines dense vector search, BM25 text retrieval, asynchronous ingestion pipelines, and a segment-based storage lifecycle into a unified architecture. ### Key Features - **Memory Control**: Disk-backed vector storage using mmap, with configurable memory budgeting and eviction policies to reduce RAM pressure. - **Async Ingestion**: Queue-based ingestion with explicit backpressure handling and stable performance under sustained write load. - **Segment Lifecycle**: Explicit lifecycle management: growing → sealed → compacted, with background compaction and improved long-running stability. - **Hybrid Retrieval**: Combined vector similarity and BM25 keyword search, with Reciprocal Rank Fusion (RRF) and deterministic result merging. - **gRPC-First API**: `proto/barq.proto` is the canonical API contract, with aligned SDKs to gRPC, and REST maintained for compatibility. ### Why Barq DB Barq-DB is designed as a retrieval system rather than a standalone vector store, enabling better control over performance, memory usage, and long-running stability. ### Key Highlights (v2) - **Memory Control**: Disk-backed vector storage using mmap, with configurable memory budgeting and eviction policies. - **Async Ingestion**: Queue-based ingestion with explicit backpressure handling and stable performance under sustained write load. - **Segment Lifecycle**: Explicit lifecycle management: growing → sealed → compacted, with background compaction and improved long-running stability. - **Hybrid Retrieval**: Combined vector similarity and BM2

### Overview: SelfElicit is an inference-time context augmentation framework designed to enhance the accuracy and relevance of language models by automatically highlighting key evidence from the context. This approach leverages pre-trained language models' attention patterns to focus on relevant information, making it model-agnostic, computationally efficient, and suitable for various LLMs. ### Key features: - **Automatic Evidence Highlighting**: Uses attention patterns in pre-trained language models to identify crucial evidence within the context, helping the model focus on relevant information. - **Model-Agnostic Booster**: Boosts the performance of various families of LLMs with different sizes and architectures, tested with Llama3.1, Mistral, and Qwen2.5. - **Light-weight Inference-Time Augmentation**: Applies evidence selection and context highlighting during inference without requiring iterative prompting or additional training. ### Tech stack: Jupyter Notebook ### Why it matters: SelfElicit enhances the accuracy and relevance of language models by focusing on relevant information, making it a valuable tool for researchers and developers looking to improve the performance and quality of their models.

### Overview: Emuru is an official PyTorch implementation for "Zero-Shot Styled Text Image Generation, but Make It Autoregressive". It integrates a T5-based decoder with a Variational Autoencoder (VAE) to generate new images conditioned on text and style images. The code is tested on Python 3.11.13, CUDA 12.8, and PyTorch 2.7.1, using a NVIDIA RTX 4090 GPU. ### Key Features: - **Conditional Image Generation**: Emuru allows users to generate new images by combining textual prompts (e.g., style text, generation text) and style images. - **Autoregressive Architecture**: The model is autoregressive, meaning it generates the image step by step, ensuring smooth transitions. - **Conditional Text**: Supports text conditioning, allowing for more nuanced image generation. - **Pretrained Models**: The code provides all intermediate artifacts, including the VAE and T5 decoder, in the Releases section. ### Tech Stack: Emuru is built using the following technologies: - **Python**: Version 3.11.13 - **CUDA**: Version 12.8 - **PyTorch**: Version 2.7.1 - **NVIDIA RTX 4090 GPU**: Used for training ### Why It Matters: Emuru is a significant contribution to the field of text-to-image generation, offering a more flexible and autoregressive approach to image synthesis. This technology can be applied across various domains, including

### Overview: ModalFormer is a state-of-the-art multimodal transformer designed specifically for low-light image enhancement (LLIE). It leverages nine auxiliary modalities to achieve outstanding performance, making it the first large-scale framework for this task. ### Key features: - **Cross-modal Transformer (CM-T)**: A specialized transformer designed to enhance corrupted images by integrating multimodal information. - **Cross-modal Multi-headed Self-Attention (CM-MSA)**: A novel mechanism that fuses RGB data with modality-specific features, including deep feature embeddings, segmentation information, geometric cues, and color information. - **Multiple Auxiliary Subnetworks**: Dedicated to reconstructing multimodal features, enhancing the overall performance of the model. ### Tech stack: Python ### Why it matters: ModalFormer stands out by addressing the fundamental yet challenging task of low-light image enhancement by fully exploiting multiple visual modalities. Its state-of-the-art performance is demonstrated across various benchmark datasets, making it a valuable tool for researchers and practitioners in the field of low-light image processing and enhancement.