Projects

Overview: Fine-tuning of LiquidAI LFM2-VL vision-language models on food image classification using LoRA. Key features: - 450M Base Model - 450M Fine-tuned Model - 1.6B Base Model - 1.6B Fine-tuned Model Tech stack: Jupyter Notebook Why it matters: This project provides a fine-tuning solution for LiquidAI LFM2-VL models on food image classification, showcasing improvements in model accuracy through fine-tuning.

Overview: LyricDiffusion is a versatile application that transforms song lyrics into visually stunning images and then those images can be combined into a slideshow. Key features: - Genius API Integration: Automatically fetches lyrics for the specified artist and song. - Customizable Inference: Choose the number of inference steps (1-100) to fine-tune the generated images. - Extra Prompt: Enhance the generated images with an extra prompt, enabling specific styles (e.g., dark ambience, extremely detailed). - Model Selection: Select a language model from Hugging Face's collection (e.g., dataautogpt3/OpenDalleV1.1). - Skip Irrelevant Verses: Exclude empty or non-lyrical verses fetched from the Genius API. - Output Path Configuration: Change the output path where the images and videos will be saved. - Video FPS: Choose how fast the slideshow should move. - Video Format: Choose in which format the video should be generated. Tech stack: Python Why it matters: LyricDiffusion is a powerful tool for visualizing song lyrics, allowing users to create stunning images and videos that can be customized to their preferences. The integration with the Genius API makes it easy to generate lyrics for any artist and song, while the customizable inference and extra prompt features enable users to fine-tune the generated images to their liking. The versatility of the application makes it a valuable tool for anyone interested in visualizing song lyrics in a creative way.

Overview: This project aims to enhance the interaction between a large language model and synthetic health data in the FHIR format, contributing to the LabLabAI AI Challenge Hackathon in October 2023. Key features: - User interface for displaying synthetic health data - Integration with a language model (GPT-3.5, GPT-4) for enhanced interaction - Connection between synthetic data and the language model - Definition and documentation of workflows and queries Tech stack: Jupyter Notebook Why it matters: This project demonstrates the integration of AI models with medical data, showcasing the potential of using AI in healthcare. Repository: bsenst/llm-enhanced-ehr Primary language: Jupyter Notebook GitHub stars: 8 GitHub description: Contribution to the LabLabAI AI Challenge Hackathon October 2023 [Project Team Page](https://lablab.ai/event/ai-challenge-with-gpt-3-5-codex-dall-e-and-whisper-api/fritzlabs) & [Project Submission Page](https://lablab.ai/event/ai-challenge-with-gpt-3-5-codex-dall-e-and-whisper-api/fritzlabs/llm-enhanced-medical-notes)  Run

Overview: This repository contains the training code for a beginner-friendly tutorial on fine-tuning image classification models using PyTorch and the timm library for a hand gesture recognizer. Key features: - Fine-tuning image classification models with PyTorch and the timm library - Creating a hand gesture recognizer - Step-by-step guide accessible to beginners Tech stack: Jupyter Notebook Why it matters: This tutorial provides a practical, accessible way for beginners to learn and apply deep learning concepts, specifically focusing on image classification and fine-tuning PyTorch models. Repository: cj-mills/pytorch-timm-gesture-recognition-tutorial-code Primary language: Jupyter Notebook GitHub stars: 8 GitHub description: Contains the training code for a beginner-friendly tutorial titled "Fine-Tuning Image Classifiers with PyTorch and the timm library for Beginners". README excerpt: This tutorial provides a step-by-step guide on fine-tuning image classification models using PyTorch and the timm library to create a hand gesture recognizer. The content is structured to be accessible even to beginners in deep learning.

### Overview: This repository is a comprehensive collection of AI logo generator prompts, workflows, and branding resources designed to help creators and entrepreneurs explore logo concepts, brand identity ideas, and visual style directions using artificial intelligence. ### Key Features: - AI logo generator prompts - AI logo maker workflows - AI logo creator ideas - AI logo maker examples - Brand identity prompts - Icon concepts - Style directions - Color palette workflows - Creative branding resources ### Tech Stack: Python and AI models. ### Why it Matters: This repository provides a practical and SEO-focused toolkit for founders, creators, marketers, designers, indie makers, agencies, e-commerce sellers, and small businesses looking to leverage AI for fast ideation in logo and visual identity design. It helps in exploring logo directions, symbols, icons, and style systems without replacing final brand design judgment or legal review.

### Overview: This repository provides a comprehensive guide to creating free AI image prompts, workflows, and resources for beginners and those looking to explore AI art generation without incurring high costs. ### Key Features: - **Free AI Image Generator Prompts**: Examples and patterns for creating AI images without spending money. - **Text-to-Image Workflows**: Prompts for generating images from text, suitable for beginners. - **AI Art Prompts**: Creative prompts for generating AI art. - **Realistic Image Prompts**: Prompts that are realistic and practical for various use cases. - **Product Image Prompts**: Prompts for creating product images. - **Profile Picture Prompts**: Prompts for creating profile pictures. - **YouTube Thumbnail Prompts**: Prompts for creating thumbnail images for YouTube videos. - **Logo Concept Prompts**: Prompts for creating logo concepts. - **Print-on-Demand Prompts**: Prompts for creating images that can be printed on demand. - **Image Editing and Upscaling Workflow Notes**: Tips and guidelines for editing and upscaling generated images. - **Prompt Formulas**: Formulas to reduce failed generations. - **Honest Guidance**: Honest advice about free limits and realistic expectations. - **Quality Checklists**: Checklists to ensure the quality of generated images. - **Safety, Copyright, and Commercial-Use Notes**: Important information about safety, copyright, and commercial use. - **Links to Relevant Cliprise Pages, Tools, and Guides**: Access to additional resources for further learning. ### Tech Stack: None ### Why It Matters: This repository is essential

Overview: This repository is a comprehensive resource for Seedream 5 prompts, Seedream prompts, Seedream 5 image prompts, AI image generation prompts, product photography prompts, ad creative templates, image editing ideas, and Cliprise resources. It serves creators, marketers, founders, agencies, e-commerce teams, app builders, musicians, educators, and visual storytellers looking for better prompts, stronger image outputs, and reusable creative assets. Key features: - Seedream 5 prompts - Seedream prompts - Seedream 5 image prompts - Seedream 5 product photography prompts - Seedream 5 poster prompts - Seedream 5 ad creative prompts - Seedream 5 social media image prompts - Seedream 5 editorial prompts - Seedream 5 character prompts - Seedream 5 concept art prompts - Seedream 5 text-to-image prompts - Seedream 5 image editing prompts - AI image generation prompts - AI commercial image prompts - AI product image prompts - AI poster design prompts - AI source-frame prompts Tech stack: None specified. Why it matters: This repository provides a practical, SEO-friendly hub for creators seeking better prompts, stronger image outputs, and reusable creative assets in the Seedream 5 and AI image generation workflows.

Overview: The A-PROOF/ZonMw project aims to develop classifiers that identify the functioning level of a patient from free-text clinical notes in Dutch, focusing on 9 International Classification of Functioning, Disability and Health (ICF) domains relevant to recovery from COVID-19. Key features: - Multi-label classification for detecting 9 ICF domains - Regression models for assigning levels of functioning per domain - Comprehensive data processing and annotation scripts - Machine learning evaluation and pipeline documentation Tech stack: Jupyter Notebook Why it matters: This project contributes to improving patient recovery assessment through automated classification, leveraging annotated data and machine learning techniques to enhance clinical decision-making. Repository: cltl/a-proof-zonmw GitHub stars: 8 GitHub description: Detecting the functioning level of a patient from a free-text clinical note in Dutch.

### Overview: Midjourney Prompts: 8 Tested Recipes With Outputs <!-- conv-kit:v1 --> <p align="center"> <img src="assets/badges/price.svg" alt="observed unit price"> <img src="assets/badges/billing.svg" alt="billing model"> <img src="assets/badges/compat.svg" alt="OpenAI-compatible endpoint"> </p> <p align="center"> <img src="assets/01-product-hero-perfume.jpg" width="820" alt="Sample render via the APIMart image API"> </p> > **from $0.045 per image** on the APIMart route — satu endpoint yang kompatibel dengan OpenAI di `https://api.apimart.ai/v1`, tanpa paket bulanan. *(diamati 2026-10-12)* **[Get an API key](https://go.apimart.ai/k-406e37)** · **[Live pricing](https://go.apimart.ai/k-88bcdf)** · **[Model page](https://go.apimart.ai/k-ec2226)** · [⚡ 60-second quickstart](#quickstart) **Alasan tim memakai Midjourney (`midjourney`) lewat APIMart** - **Satu key untuk seluruh katalog.** base URL `https://api.apimart.ai/v1` dan header `

### Overview: The CyborgDB Python SDK is a Python client designed for the CyborgDB encrypted vector database, providing a self-hosted solution for secure vector search. This SDK encrypts data at rest, ensuring that Cyborg Inc. never receives keys, vectors, or queries, making it ideal for protecting sensitive data. ### Key Features: - **Encryption-in-use**: Search runs over ciphertext, only the final top-k candidates are decrypted, ensuring data security. - **Encrypted ANN**: Disk-backed encrypted DiskIVF index, providing high performance. - **Filters on encrypted metadata**: Combine vector similarity with equality and range predicates in a single request. - **BYOK / HYOK**: Wrap per-index keys with AWS KMS or AWS Secrets Manager, or hold the key client-side. - **Per-tenant key isolation**: Per-index, per-user keys with cryptographic RBAC; keys can be revoked. - **Pythonic API**: Familiar client/index interface that integrates seamlessly with existing Python AI workflows. ### Tech Stack: The CyborgDB Python SDK is built on Python, leveraging the `cyborgdb-service` Docker image for self-hosted deployment. ### Why It Matters: This SDK offers a secure, self-hosted solution for vector search, protecting sensitive data and ensuring compliance with encryption-in-use practices. The Pythonic API makes it easy to integrate with existing AI workflows, making it a valuable tool for developers looking to enhance their data processing capabilities.

Overview: A privacy-focused document retrieval and chat app for macOS that indexes local documents and allows direct, offline-based search and chat. Key features: - Offline document indexing and search - Secure, privacy-first retrieval and augmented generation - No cloud storage or tracking Tech stack: Swift Why it matters: This app provides a secure and privacy-focused alternative to cloud-based document management systems, ensuring users can access their documents offline without compromising their privacy.

Overview: Seedream 4.5 API (`seedream-4-5`) provides a cost-effective OpenAI-compatible endpoint for generating high-quality images, videos, and text-based content. With a single API call, users can access a catalog of over 300 models from ByteDance, including text-to-image generation, style fusion, and unified editing. Key features: - **Model Id and Routes**: The API supports the `seedream-4-5` model id, with pricing based on the number of delivered images. - **Pricing Model**: Prices are effective per delivered image, with a default rate of $0.0325 and a maximum of $1 per task. - **Billing Model**: The pricing is billed on the cost of the generated images, and tasks can be resumed or recharged. - **Assynchronous**: The API is designed to be asynchronous, allowing users to track task progress with `GET /v1/tasks/{task_id}`. Tech stack: Python Why it matters: Seedream 4.5 API offers a cost-effective alternative to OpenAI's API, providing high-quality image generation at a lower price point. Its catalog of models and features make it a valuable tool for developers looking to integrate AI into their projects without the high costs of a dedicated API key. Repository: d4qb-s4y087m/seedream-4-5-api-pt Primary language: Python GitHub stars: 8 GitHub description: Seedream 4.5 API (`seedream-4-5`):

### Overview: This repository provides a comprehensive toolkit for fine-tuning Gemma 4 on Amazon SageMaker across five self-contained tracks: JSON extraction, intent classification, summarization, domain QA, and multimodal receipt parsing. It includes synthetic data, LoRA/QLoRA SFT with optional GRPO, vLLM/SGLang/LMI serving, held-out evaluation, and a Bedrock Claude agentic loop. ### Key features: - Fine-tuning Gemma 4 on Amazon SageMaker - Covers synthetic data, LoRA/QLoRA SFT with optional GRPO, vLLM/SGLang/LMI serving, held-out evaluation, and a Bedrock Claude agentic loop - Tracks: JSON extraction, intent classification, summarization, domain QA, and multimodal receipt parsing - Supports both local and cloud execution - Includes detailed documentation and a guidebook for beginners ### Tech stack: Python ### Why it matters: This toolkit provides a practical, hands-on approach to fine-tuning and deploying a model on SageMaker, making it ideal for developers looking to integrate their models into production environments. The included tracks and examples cover a wide range of use cases, from basic JSON extraction to advanced multimodal receipt parsing, ensuring that users can quickly get started and build their models effectively.

Overview: Qube is a privacy-first desktop AI assistant that unifies personal knowledge management with institutional research sources and inspectable, citation-grounded answers on the user's hardware. Key features: - Composer & `@` routing - Library - Memory Manager - Live Sources Tech stack: Python Why it matters: Qube offers complete privacy and sovereignty, grounded answers, and transparent local telemetry, making it ideal for sensitive work on one's own hardware.

Overview: nxm-memory is a local memory and search engine designed for AI assistants, indexing code and documents on your machine for offline, private, and fast access. It supports macOS arm64 and Linux x86_64, and integrates with MCP-compatible tools like Claude Code, Opencode, Pi, and others. Key features: - Indexes code and documents on your machine. - Queryable via MCP. - Compresses source code into structural maps and shell output for faster processing. - Retrieves only relevant chunks from documents for efficient search. - Supports private, offline, and fast access. - Configured exactly like any other MCP server. Tech stack: Shell Why it matters: It reduces the number of tokens sent to the model, making it more cost-effective and better suited for context-rich environments. GitHub stars: 7 Repository: [nxm-memory](https://github.com/dangranaz/nxm-memory)

### Overview: SuperSLM is a deterministic small-language-model inference runtime designed specifically for real-time games. It loads a converted, quantized `.sslm` artifact and runs it with integer-only arithmetic, ensuring identical output across CPUs, GPUs, compilers, and vendors. ### Key Features: - **Deterministic Output**: Ensures identical output across different platforms and configurations. - **Bit-Identical Output**: The same model, prompt, and decoding configuration produce bit-identical tokens every time. - **Frame Budget Efficiency**: The engine can slice a decode step across a caller-chosen number of transformer layers per call, spreading work across frames for better GPU utilization. ### Tech Stack: C++ ### Why it Matters: Being deterministic is crucial for real-time games, as it allows the engine to throttle inference to fit the available GPU headroom without changing the model's output. ### Repository: dansupergameprogrammer/SuperSLM ### GitHub Stars: 8

### Overview Ink Traces WebUI is an open-source web UI designed for Gemini, Seedream, and Seedance text-to-image and image-to-image generation. It provides a unified workflow for prompt-driven image and video creation, supports multiple model providers, and stores history and data outside of Git. ### Key Features - **Prompt-driven Image and Video Creation**: Combines a React/Vite frontend with a Flask backend, supports multiple model providers, and stores generation history in SQLite. - **Image Generation**: Text-to-image and image-to-image through one unified workflow. - **Video Generation**: Seedance 2.0/2.5 keyframe and multimodal reference workflows with image, video, and audio inputs. - **Providers**: Google Gemini API Nano, Banana 2.1, and BytePlus Ark for images; BytePlus Ark Seedance 2.0/2.5 for video. - **Prompt Workflow**: Prompt Vault, fullscreen editor, multi-tab workspaces, reusable saved prompts. - **Runtime History**: SQLite task queue for image/video results, local file recovery, task restore. - **Controls**: Aspect ratio, resolution, thinking level, Google Search grounding, chat mode. - **Privacy**: Real `config.json`, Prompt Vault data, logs, generated outputs, uploads, and DB files are ignored. ### Tech Stack - **Frontend**: React 18, Vite, Tailwind CSS, Framer Motion, Lucide Icons. - **Backend**: Python Flask, Flask-CORS, Pillow, Requests. - **Storage**: SQLite task database

Overview: This project is an AI-powered Multi-Agent Research Assistant built with advanced technologies like Generative AI, Agentic AI, LangGraph, RAG, Persistent Memory, FAISS, and FastAPI. It supports multi-format uploads, intelligent PDF analysis, and expert-like Q&A via Researcher, Summarizer, Critic, and Editor agents, offering deep, contextual, and interactive research insights. Key features: - Multi-Agent Workflow: Research Agent, Summarizer Agent, Critic Agent, and Editor Agent work together to simulate real research processes. - Multi-Document Support: Allows uploading and querying multiple PDFs, DOCX, or TXT files. - Conversation Memory: SQLite-backed memory for contextual, follow-up queries. - Chunking & Embeddings: Splits documents into chunks and embeds them using OpenAI models. - FAISS Vector Search: High-speed semantic retrieval of embedded text chunks. - LangGraph Orchestration: Structured multi-agent pipeline with conditional routing. - Modern UI: Responsive frontend built with Next.js and TailwindCSS. Tech stack: Python, FastAPI, Next.js, TailwindCSS, SQLite, FAISS, OpenAI models, LangGraph, Persistent Memory, RAG, FastAPI. Why it matters: This project offers a comprehensive solution for researchers, providing a multi-agent system that enhances the efficiency and quality of research by offering deep, contextual, and interactive insights.

Overview: This repository contains 8 tested FLUX 2 Pro prompts with example renders, aspect ratios, and the cost per image on the `flux-2-pro` model. The prompts are designed to guide the model towards more accurate and meaningful results. Key features: - 8 FLUX 2 Pro prompts - Example renders, aspect ratios, and call costs - Pricing: $0.024 per image (1MP $0.024, 2MP $0.036, 3MP $0.048) Tech stack: Python Why it matters: The prompts are designed to help developers and designers create more accurate and visually appealing images, reducing the need for manual adjustments and ensuring consistent quality across different projects.

### Overview: Profile-guided layout optimization for MoE model files, reducing disk reads by 2.23x compared to the stock layout. ### Key Features: - **Rewrite Layout**: Reorders expert slices based on real routing traces to improve sequential access. - **Trace**: Logs which experts each token selects. - **Rewrite**: Reorders expert slices by measured co-activation. - **Rewrite**: Uses interleave layout for contiguous per expert blocks. - **Prefetch**: Explicitly reads selected slices to reduce page faults. ### Tech Stack: Python ### Why It Matters: Profile-guided optimization for MoE model files, reducing disk reads by 2.23x, improving inference performance.