AI

LLM Chat RAG

LLM Chat RAG - CLI Interface Overview The LLM Chat RAG (Retriever-Augmented Generation) is a Command Line Interface (CLI) tool that allows you to interact with documents stored in a ChromaDB database using OpenAI’s GPT-4o-mini model. This system provides a way to query documents for relevant information, retrieve context, and generate AI-driven responses based on that context. Features Retriever-Augmented Generation (RAG): Retrieve context from a database and use it to generate more accurate and informative responses. OpenAI GPT-4o-mini Model: Uses OpenAI’s GPT-4o-mini to generate answers based on the retrieved context. ChromaDB Integration: Uses ChromaDB for efficient document retrieval and context management. Command-line interface: Interact with the system through an intuitive CLI, allowing commands such as /help, /exit, and /sources. Prerequisites Python 3.x: Ensure Python 3.6 or later is installed. OpenAI API Key: You must have an OpenAI API key to use the GPT model. ChromaDB: A local or cloud-based ChromaDB instance for document storage and retrieval. Setup 1. Install Dependencies Install required dependencies via pip:

March 13, 2025 Read

Database Embedding

Markdown to ChromaDB Indexer This tool indexes markdown files into ChromaDB for efficient semantic search capabilities, with support for both default embeddings and OpenAI’s text-embedding-3-small model for enhanced search quality. Features Flexible embedding options: Default ChromaDB embeddings (no API key required) Optional OpenAI text-embedding-3-small model for enhanced quality Recursively processes markdown files in a directory Intelligent text chunking with configurable size and overlap Sentence-aware splitting to maintain context Extracts and preserves frontmatter metadata Converts markdown to searchable text Stores documents with their metadata in ChromaDB Supports semantic search queries Batch processing for large datasets Installation Install the required dependencies: pip install -r requirements.txt (Optional) Set up OpenAI embeddings: Create a .env file with your OpenAI API key: OPENAI_API_KEY=your_api_key_here Usage To index your markdown files:

March 6, 2025 Read

Subtitle Generator

Subtitle Generator This project provides a web application that automatically generates subtitles (in SRT or VTT format) from audio files. It leverages the power of OpenAI’s Whisper speech recognition model for transcription and pyannote.audio for speaker diarization, all wrapped in a user-friendly FastAPI web interface. Features Web Interface: A clean, simple HTML interface for uploading audio files. No complex command-line usage is required. Multiple Audio Formats: Supports .wav, .mp3, .m4a, and .flac files. Automatic conversion to WAV is handled internally. Speaker Diarization: Identifies different speakers in the audio and includes speaker labels in the generated subtitles. Subtitle Formats: Generates subtitles in either .srt (SubRip) or .vtt (WebVTT) format, selectable via the web interface. Fast and Efficient: Utilizes a dynamically selected Whisper model (from tiny to large) based on your system’s available RAM and VRAM (GPU memory), optimizing for performance and accuracy. Easy Deployment: Can be run directly with Uvicorn or easily deployed using Docker and Docker Compose. Well-Defined API: A single /upload-audio/ endpoint handles file uploads and subtitle generation, with clear request and response formats. File Size Limit: The maximum upload file is limited to 50MB. Requirements Before you get started, make sure you have the following:

February 20, 2025 Read

Object Tracker

The statement Long-Term Visual Object Tracking Benchmark is a dataset to benchmark object tracking algorithms. We were asked to implement four different algorithms. Two using tracking techniques, and two using recognition. For each technique, we had to implement one from scratch (an algorithm made for us) and another one extracted from the internet. Given a series of frames and a bounding box corresponding to the object to track, the algorithm should follow the object. To evaluate the precision, we used the overlapping ratio between the computed bounding box and the optimal bounding box (which was given via a txt file).

June 23, 2023 Read

Royal Hackaway v6

About Royal Hackaway Royal Hackaway is Royal Holloway’s annual hackathon. Open to students from all around the world, in this two-day hackathon you’ll work as a team to build a project. You’ll pick up new skills, try out new tech, and meet new people. This is our sixth time running this event, and we’re excited to create a more inclusive and rewarding event for all. You can join us in-person, or online!

February 7, 2023 Read

Mars Planification

The statement We are in the year 2075 and the first colonizers of Mars have established a network of bases throughout the geography of the planet that can be of two types, settlements (where the settlers live) and warehouses (where supplies arrive from the Earth). Moving between these bases is done by rovers that can move supplies and specialized personnel. It is not possible to always go directly from one base to another, we have a map (connected graph) that tells us which movements are possible.

January 14, 2023 Read

Workout Planner

The statement The progressive aging of the population is showing the importance of maintaining a set of healthy habits and promote physical exercise throughout life. When you reach a certain age it is necessary to have help to determine what types of exercises and activities are the most appropriate taking into account all factors that may affect the performance of those activities. Basically, we had to make a CLIPS script that after given certain information about a person, we had to output an exercise planning suitable for him/her. We were given a bunch of documents about exercises for elder people where we had to extract the information.

December 19, 2022 Read

Grid Power Optimization

The statement Given a set of Power plants and Clients, find an assignation for each Client where the total benefit is the maximum. Each power plant is represented by coordinates in a map, the total production and the cost when it’s stopped and running. Each client is represented by coordinates in a map, the power consumed, the price it pays for each Mw and a compensation in the case he doesn’t get power. Also, power gets drained if the distance between a power plant and a client is farther than a certain value.

October 24, 2022 Read