Architecture & Design

DocAgent employs a sophisticated multi-agent architecture where specialized agents handle different aspects of document processing:

• Ingestion Agent: Handles document parsing and pre-processing
• Retrieval Agent: Performs semantic search using FAISS
• LLM Response Agent: Generates responses using Llama 3.1 8b
• Coordinator Agent: Orchestrates communication between agents

Tools & Tech Stack

Built with LangChain, HuggingFace Inference Client, Llama 3.1 8b, FAISS, and Gradio. The system leverages Retrieval-Augmented Generation (RAG) with a Model Context Protocol (MCP) for coordinated multi-format document QA.

Key Achievements

• Achieved Word Error Rate (WER) of 22
• Preprocessed and augmented dataset for accent diversity
• Used Hugging Face for training and versioning

Training and evaluation data

The model was fine-tuned on a custom web-scraped dataset called "TamilVoiceCorpus" (available at ragunath-ravi/TamilVoiceCorpus on Hugging Face). This dataset consists of Tamil speech audio paired with accurate transcriptions. For training, 12,000 samples were used from the dataset, while 3,000 samples were used for evaluation. The audio was processed by resampling from 48kHz to 16kHz to match Whisper's requirements. The dataset includes a diverse range of Tamil speech samples, which helps the model generalize across different speakers, accents, and content types.

Dataset Structure

• PureVox

The PureVox subset contains cleaner audio with minimal background noise, making it ideal for baseline ASR model training and evaluation.

• AmbientVox

The AmbientVox subset includes audio with background noise, simulating real-world conditions for more robust ASR model training.

Info

This is a dataset for Automatic Speech Recognition (ASR) focused on conversational Tamil, collected from various public sources on the web. Each sample is a short audio clip (averaging 10 seconds) paired with its corresponding transcription.

Dataset Summary

• Split: Chain of Thought (CoT) only
• Total Rows: 3,201,061
• Data Format: Plain text, LLM-tokenizer friendly
• Primary Use Case: Training and evaluating LLMs on complex mathematical reasoning tasks using chain-of-thought prompting

Source

• Original Dataset: NVIDIA Open Math Reasoning
• This Version: Chain-of-Thought split, extracted and preprocessed for LLM use (cleaned, formatted, deduplicated where necessary)

Task Management

• Create, edit, and delete tasks with rich details.
• Quick add feature for rapid task entry.
• Mark tasks as complete with one click.
• Organize tasks by categories and priorities..
• Add detailed notes to each task.

Technical Benefits

• Lightweight and efficient: Optimized for performance.
• Intuitive UI: Designed for effortless usability.
• Robust data handling: Ensures data safety and persistence.

Project Overview

The system processes video frames to detect space debris, predict its future position, and visualize the trajectory. It uses YOLOv5 for real-time object detection and a Kalman Filter for trajectory prediction.

Features

• Real-Time Detection: YOLOv5 detects debris in video frames. Kalman Filter
• Prediction: Predicts the debris's future position based on previous frame positions.
• Visual Tracking: Draws bounding boxes and predicted trajectory on each frame.
• Video Output: Saves a new video file with tracked debris and predictions.

Music Streaming

• Search for music on YTube,Ytmusic.
• Stream music directly from YTube.
• Display related songs based on the current track.

Offline Playback

• Download songs for offline listening.
• Manage downloaded songs and playlists.
• Play music in the background while using other apps.

Model Training & Evaluation

Trained the BLIP model on a high-performance GPU cluster with mixed precision, using a custom dataset of histopathology images.

Training Details

• Dataset: Histopathology images with corresponding captions
• Epochs: 10
• Base Model: Salesforce/blip-image-captioning-base
• Batch Size: 4
• Learning Rate: 5e-5
• Gradient Accumulation: 2