Projects

Overview: This project is an LLM-powered tool that converts high-level ethical requirements into actionable ethical user stories, using HuggingFace, retrieval-augmented generation (RAG), ChromaDB, and LangChain to bridge the gap between abstract ethical principles and practical implementation. Key features: - Ethical user stories generation - Update on RAG database locally Tech stack: Jupyter Notebook Why it matters: This tool ensures that ethical considerations are seamlessly integrated into software development processes, making it essential for ethical software development practices. Repository: joaorossi15/requirements-to-us Primary language: Jupyter Notebook GitHub stars: 4 GitHub description: Ethical requirements to ethical user stories README excerpt: This project is an LLM-powered tool that converts high-level ethical requirements into actionable ethical user stories, using HuggingFace, retrieval-augmented generation (RAG), ChromaDB, and LangChain to bridge the gap between abstract ethical principles and practical implementation.

Overview: DIRSM is a project that focuses on fine-tuning ConvNets for disaster image retrieval from social media, utilizing various architectures such as AlexNet, VGG16, GoogleNet, ResNets, and DenseNets. Key features: Fine-tuning of ConvNets using TensorFlow, including AlexNet, VGG16, GoogleNet, ResNets, and DenseNets, with algorithms for ranking generation and evaluation. Tech stack: Python, TensorFlow, Python-torchfile, tensorflow-resnet, caffe-tensorflow. Why it matters: This project is crucial for improving disaster image retrieval from social media, enhancing emergency response capabilities, and aiding in the early detection of disasters.

### Overview: ResearchRadar: An AI Literature Discovery and Classification Pipeline ResearchRadar is an end-to-end pipeline designed to identify specialized research papers through automated classification. This project showcases how researchers can build an automated pipeline using ZenML's end-to-end ML pipelines to train models that classify papers relevant to their specific subfields. The pipeline includes data ingestion, model training, evaluation, and deployment, providing a comprehensive solution for researchers to stay current with AI literature without manual effort. ### Key Features: - **Datasets for Evaluation, Augmentation, and Training**: A composite dataset of manually labeled positive and negative examples, as well as non-classified (unlabeled) examples for augmentation. - **Multi-environment Training Support**: The pipeline can be run locally or on remote infrastructure via step operators. - **Comprehensive Metrics**: Track accuracy, F1 score, latency, memory usage, and cost analysis. - **Comparative Analysis**: Benchmark the fine-tuned model against Claude Haiku. ### Tech Stack: Python ### Why It Matters: This framework is ideal for academic labs, industry research groups, and individual AI researchers who need to maintain comprehensive awareness of relevant publications without spending hours each week manually sifting through papers. By deploying the trained model to Hugging Face, teams can create shared resources that help all members stay current with literature relevant to their collective research agenda, streamlining collaboration and reducing duplicate effort in literature reviews. ### Repository: marwan37/research-radar ### GitHub Stars: 4 ### GitHub Description: End-to-end pipeline that identifies specialized research papers

Overview: This project serves as a layer for large machine learning models on Apache Flink, providing a scalable and efficient solution for deploying and processing machine learning models in the cloud. Key features: - Integration with Apache Flink for processing and managing large datasets - Support for serving machine learning models in real-time - Scalability to handle varying loads and data volumes Tech stack: Java Why it matters: This project is crucial for organizations looking to deploy and manage machine learning models in the cloud, ensuring that they can leverage the power of Flink for processing and serving large datasets efficiently.

Overview: This Go library provides a simple and efficient interface for interacting with Modzy's Machine Learning Operations (MLOps) Platform, enabling developers to query models, submit inference jobs, and retrieve results directly in their editor. Key features: - Querying models - Submitting inference jobs - Retrieving results directly in the editor Tech stack: Go Why it matters: This library simplifies the integration of Modzy's MLOps Platform into existing projects, allowing developers to leverage pre-trained and re-trained AI models for faster development cycles and improved accuracy.

Overview: This project explores the visualization of textual information using deep neural models trained for image processing, focusing on text classification and fractality analysis. Key features: - Text classification using pre-trained deep neural models for image classification - Fractality analysis of converted text to images using deep neural models Tech stack: Python, PyTorch, TensorFlow Why it matters: This project demonstrates innovative ways to visualize textual data, making complex information more accessible and interpretable.

### Overview: Mesh Deck is a command-line tool designed to plan, order, and run multiple local AI services on NVIDIA GPUs. It ensures that each service starts in the correct order, preventing memory issues that can occur when services compete for GPU resources. ### Key Features: - **Memory Checks:** Simulates GPU memory usage for each service in the order they will run, identifying which services will not fit on the GPU. - **Ordering Services:** Ensures that services start in the correct order, avoiding conflicts for GPU resources. - **Single Log File per Service:** Each service has its own log file, making it easy to track and debug memory issues. - **Real Memory Reporting:** After each service starts, it reports the amount of GPU memory added by that service. - **Memory Holder Identification:** Reveals which service is holding the GPU memory, helping to pinpoint the root cause of memory issues. - **Clean Shutdown:** Stops all services in reverse order, providing a clear and complete shutdown. - **Readable in Terminal:** Uses color, tables, and memory bars to make the output clear and easy to read, even when piping or using `NO_COLOR`. ### Tech Stack: - Python 3.10+ ### Why It Matters: If you run several GPU services on the same machine, you might encounter issues where the second service crashes due to insufficient GPU memory. Mesh Deck helps you plan and order your services in the right sequence, preventing such issues and making troubleshooting easier.

Overview: ViAG is a Vietnamese answer generation framework that fine-tunes encoder-decoder models on Vietnamese question-answering tasks. It supports local CSV datasets and provides comprehensive evaluation metrics. Key features: - Fine-tune pre-trained encoder-decoder models for answer generation - Support for local CSV datasets - Comprehensive evaluation metrics (ROUGE, BLEU, METEOR, BERTScore) - Command-line interface for training and evaluation - Weights & Biases integration for experiment tracking - Modular and extensible codebase Tech stack: Python Why it matters: ViAG simplifies the process of fine-tuning Vietnamese answer generation models, making it accessible for developers and researchers alike.

Overview: This repository contains my AI projects, showcasing my latest advancements in the field of artificial intelligence. Key features: - Implementing advanced natural language processing models - Developing chatbots for customer service - Creating image generation algorithms Tech stack: Jupyter Notebook Why it matters: These projects demonstrate my ability to tackle complex AI problems and contribute to the field through practical applications.

Overview: A versatile AI Admin and Assistance Bot designed to assist users and perform administrative tasks. Key features: - AI-powered chatbot capabilities - Integration with various user management systems - Customizable responses and intents - Integration with multiple databases Tech stack: JavaScript Why it matters: This bot streamlines administrative processes, enhancing user experience and operational efficiency.

### Overview: This repository contains reproducible QLoRA instruction-tuned checkpoints for multilingual Grammatical Error Correction (GEC) tasks. It includes lightweight notebooks, prompt code, and configs for fine-tuning Aya-Expanse-8B and Gemma-3-12B models. The checkpoints are optimized for state-of-the-art performance and are available on Hugging Face Hub. ### Key Features: - **QLoRA Fine-Tuning Recipes**: Reproducible QLoRA recipes for Aya-Expanse-8B and Gemma-3-12B, fine-tuned for multilingual GEC tasks. - **Minimal Edits Track**: Models optimized for minimal edits, suitable for applications where high fluency is not a priority. - **Fluency Edits Track**: Models optimized for fluency edits, providing better readability and naturalness. - **Small-LLM Pilot Track**: A small LLM pilot model optimized for Gemma-3-4B, suitable for pilot projects. ### Tech Stack: - **Jupyter Notebook**: The primary development and execution environment. ### Why It Matters: This repository provides a comprehensive solution for multilingual GEC tasks, offering high-quality checkpoints that are optimized for both fluency and minimal edits. The reproducibility ensures that researchers and practitioners can easily adapt and evaluate these models for their specific needs. ### Repository Layout: - **notebooks/**: Contains training and inference notebooks for Aya-Expanse-8B and Gemma-3-12B. - **params/**: Contains

Overview: This repository contains the code and resources for fine-tuning the Alpaca 7b model for the Sinhala language. The Alpaca 7b model is a powerful language model, and this project aims to adapt it to the Sinhala language through fine-tuning. Key features: Fine-tuning of the Alpaca 7b model for the Sinhala language, dataset translation, pre-trained model available, Colab notebook for easy experimentation. Tech stack: Python. Why it matters: This project aims to improve the performance of language models for the Sinhala language, potentially leading to better communication and understanding in the region.

Overview: Krypton is a model server built with FastAPI, designed for deploying Machine Learning and Deep Learning models. It simplifies the process of integrating ML and DL models into applications using Python scripts. Key features: - Model Server for ML and DL models - Deploy ML and DL models using simple Python scripts Tech stack: Python Why it matters: Krypton simplifies the deployment of ML and DL models, making it easier for developers to integrate these models into their applications without the need for complex infrastructure setup. Repository: saivarunk/krypton Primary language: Python GitHub stars: 4 GitHub description: Model Server for ML and DL Models built using FastAPI README excerpt:

Overview: Reasonable Doubt is a Python-based project that classifies contract clauses into LEDGAR's 100 provision types. It achieves a macro-F1 score of 0.8091 with a cost of $0.00088 per thousand clauses, offering a significant cost savings compared to other models. Key features: - Fine-tuned DeBERTa model on CPU with a cost of $0.00088 per thousand clauses. - Architecture: one tier, plus a `needs_review` flag. - Live service available at <https://reasonable-doubt-111680840326.asia-south1.run.app> with a warm-up period of 21 seconds and a cold start of 2 minutes. Tech stack: Python. Why it matters: Reasonable Doubt provides a cost-effective solution for contract clause classification, offering a 50% reduction in cost compared to frontier models while maintaining a macro-F1 score of 0.8091.

### Overview: Logistik is a machine learning logistics and serving platform designed to streamline operations and optimize performance through advanced analytics and automation. ### Key Features: - **Advanced Analytics**: Utilizes machine learning models to predict and optimize logistics processes. - **Real-time Optimization**: Implements real-time adjustments to improve efficiency and reduce costs. - **Scalability**: Designed to handle large volumes of data and traffic. - **Integration**: Seamless integration with existing logistics systems and platforms. ### Tech Stack: - **Programming Languages**: JavaScript - **Dependencies**: Node.js, Gunicorn, Consul, Ianitor ### Why It Matters: Logistik leverages machine learning to enhance logistics operations, making them more efficient and cost-effective. Its real-time optimization capabilities ensure that logistics processes are always running at peak performance, leading to better outcomes and reduced operational costs.

### Overview: Fine-tuning BERT for sequence tagging tasks often involves adding a linear layer on top of the pretrained language model. However, this approach can lead to suboptimal performance due to the absence of mechanisms for adjusting weights past the last attention operation. This repository aims to challenge the status quo by providing evidence that removing the last Transformer block can produce better initial state representations, leading to improved fine-tuning outcomes. ### Key Features: - **Removing the Last Transformer Block**: The solution elegantly removes the last Transformer block to bypass the limitations of the gradients reaching past the last attention operation. - **Empirical Evidence**: The experiment fine-tunes the model on multilingual XML-V on CoNLL 2003 and tests its performance on Afrikaans NER, demonstrating improved initial state representations. - **Training Configurations**: Multiple training configurations are tested, including fine-tuning with and without the last Transformer block, to compare the effects on performance. ### Tech Stack: The repository utilizes Jupyter Notebook as the primary language for experimentation and analysis. ### Why It Matters: Fine-tuning BERT with the last Transformer block can lead to suboptimal performance due to the absence of mechanisms for adjusting weights past the last attention operation. Removing the last Transformer block can lead to better initial state representations, resulting in improved fine-tuning outcomes.

Overview: This program is designed to help users collect, curate, and export fine-tuned data from OpenAI for fine-tuning purposes. Key features: - Supports text, image, JSON schema, function call, audio, and PDF messages. - Import OpenAI message JSON from clipboard or `.json`/`.ftproj` files. - Execute local functions with optional pre-text and store results. - Automatically upload images to a configurable server. - Configure EXIF image auto-rotation settings. - Batch-create conversations from multiple files and auto-upload created images with progress feedback. - Save/load projects locally or via a configurable cloud PHP endpoint. - Reset "New fine-tune" project with save/discard/cancel guard for unsaved changes. - Auto-save projects every 5 minutes or on conversation switches. - Cloud mode enforces URL-only images (base64 images must be uploaded first). - Export images as base64 or URLs and choose which conversations to export. - Token usage and cost estimation via a script. - Conversation auto-splitting for Reinforcement Fine-Tuning. - Optional global system message and user names per message. - Show or hide meta information and token calculations per conversation. - Optional schema-constrained assistant autocompletion for generated responses. - OpenAI-compatible API base URL mode with selectable/custom model behavior. - JSON schema array items can be moved and duplicated directly in the form UI. - Improved small-screen dialog and schema form usability. - More reliable JSON schema rendering/value restoration after reload. Tech stack: GDScript Why it matters: This program streamlines

### Overview: 本项目旨在评估语言模型在ES查询生成和数据微调方面的性能，通过导入问题集生成ES查询语句，并对生成的查询语句进行评估和微调。项目包括多个脚本和文件，涵盖了问题生成、查询语句处理、评估以及微调模型的全过程。 ### Key features: - `es_query_generate.py`: 通过导入问题集生成ES查询语句。 - `es_query_result.py`: 检测并保存成功查询的ES查询语句及其结果。 - `es_query_result_eval.py`: 对评估结果进行打分并保存打分中间文件。 - `clean_strings.py`: 删除批量生成的ES查询语句中不必要的字符。 - `finetuned_models/` 和 `finetuning_data/`: 包含微调模型的相关文档和脚本。 - `generated_queries/` 和 `generated_queries_ft/`: 保存测试的ES查询语句。 - `queries_record/` 和 `queries_record_ft/`: 保存打分结果的中间文件。 - `queries_result/` 和 `queries_result_ft/`: 保存查询结果及其打分。 - `questions/`: 包含三个问题集文件。 ### Tech stack: Python ### Why it matters: 通过评估和微调ES查询生成和数据微调的模型，该项目旨在提高模型在实际应用中的性能，特别是在处理复杂查询和生成高质量微调数据方面。这有助于提升模型在搜索和数据分析中的表现，从而

Overview: Amazon RecSys is an end-to-end recommender systems project that trains a multi-stage recommendation pipeline on Amazon review data, packages the trained artifacts into versioned serving bundles, tracks experiments with MLflow, and serves recommendations through a FastAPI web application with an analytics-focused UI. Key features: - Trains hybrid retrieval plus ranking pipelines - Ranks candidate sets with XGBoost by default, with `dlrm` as an experimental backend - Exports versioned ONNX-backed serving bundles for online inference - Serves recommendation, history, model, evaluation, and monitoring endpoints through FastAPI plus a Jinja UI - Logs training and monitoring runs to MLflow when enabled - Computes batch feature drift and concept drift from served inference logs and delayed outcomes Tech stack: Python Why it matters: This project demonstrates the importance of early model tuning and the need for a cohesive approach to training, serving, and monitoring recommender systems. It also highlights the practical considerations of operationalizing a recommender system, such as avoiding serving-time scans over full interaction tables and making readiness checks manifest-based.

### Overview: This project is a scalable image classification API built with FastAPI and ONNX, deployed on Kubernetes using Horizontal Pod Autoscaling (HPA). The API is exposed via an Ingress with a custom domain, and it includes unit tests, a hexagonal architecture, and Dockerized components. The project is designed for high performance and auto-scaling, making it ideal for real-time image classification tasks. ### Key features: - **ONNX Runtime**: No heavy PyTorch runtime, enabling faster inference. - **FastAPI**: Built with FastAPI for high-performance API endpoints. - **Kubernetes**: Deployed with HPA for auto-scaling. - **HPA**: Auto-scaling via metrics-server. - **Hexagonal Architecture**: Adapters, Core, Domain, and API. - **Dockerized**: Ready for Kubernetes deployment. - **Unit Tests**: Pytest-based, CI-ready. - **Image Uploads**: Accepts `multipart/form-data` image uploads. ### Tech stack: Python ### Why it matters: This project is designed for high-performance image classification tasks, making it suitable for real-time applications where speed and scalability are crucial. The use of FastAPI, ONNX, and Kubernetes ensures that the API is both fast and scalable, making it an excellent choice for developers looking to build efficient image classification services.