Hey there! I'm Yashaswini Ippili,

Tools Used : Python, Llama3-8b, Claude-3.5, GPT-3.5, DSPy, Firestore, ReactJS, FastAPI

Features head-to-head comparisons of LLMs where models are randomly paired and their names remain undisclosed until after the user selects the better response.

Utilizes a human vote-based Elo-rating system to evaluate and rank LLM performance on prompt generation quality, with results stored in a Firestore database.

Tools Used : Python, Llama2, Huggingface, Langchain, Transformers, Selenium, Gradio

Leveraged advanced automation techniques to seamlessly handle the processing of an extensive volume of GEHC-related PDFs. This involved automating the download, precise text extraction, merging, and meticulous cleaning of data. Through this refined dataset, facilitated the training of a sophisticated multilingual bot. This bot possesses the prowess to adeptly address inquiries spanning installations, manuals, services, and effortlessly maintains continually updated transcripts.

Achieved the position of Second Runner Up in the National Level Hackathon organized by GE Healthcare, titled Hack'E'lth'23.

Tools Used : Python, PyMySQL, Scholarly, Streamlit, SQL, Selenium

The project involves scraping Google Scholar's data for university professors. After preprocessing and cleaning, the data is stored in a MySQL database. A Streamlit front-end is developed for populating the database and displaying statistics for both individual professors and aggregated data based on specific years or combined information. It also enables detailed viewing of professors, showcasing their citations, publications, conferences, and other relevant scholarly contributions.

Tools Used : JavaScript, SQL, PHP, CSS, SCSS, Xampp, Selenium

Successfully built a full-stack web application adhering to software engineering principles and the Agile methodology. The project involved comprehensive requirements engineering, leveraging Selenium for automated testing, and utilizing tools such as SRS (Software Requirements Specification), RTM (Requirements Traceability Matrix), and Gantt charts for project management.

Tools Used : Python, PyTorch, Flan-T5, Knowledge Distillation

Novel Architecture Idea for Optimizing Large Language Models using a combination of Retrieval Augmented Generation and Multi Teacher Knowledge Distillation method.

The use of multiple teacher models allows the student model to learn a diverse range of information and patterns in order to generate the most accurate response for the user's query.

Tools Used : Python, Huggingface, Gemini 1.5 Pro, LLama3.1-8b-Instruct

A user enters a query, and a response is generated. A judge then produces N judgments using high temperature. Another judge compares two judgments at a time, performing 2*NC2 comparisons. The judgment with the most wins from these comparisons is chosen as the best judgment. This winning judgment is used to refine the initial response, aiming to produce the best possible response.

The process continues until an evaluation criterion is met, at which point the loop ends.

Tools Used : Python, Gemini 1.5 Pro, Phi3, Llama3.1-8b-Instruct

A framework designed to combine the Monte-Carlo Tree Search (MCTS) algorithm with a custom self-refine algorithm, effectively managing the exploration-exploitation trade-off to consistently find the best solution.

While the paper focuses on mathematical reasoning, this project adapts the concept to enhance the logical reasoning capabilities of an LLM along with the concept of "Meta-Judge" provided by Meta. Notably, the phi3-mini model was able to reach GPT-3.5 level logical reasoning performance.

Exploring and comparing DSPy and Conva.AI to determine the optimal AI assistant solution.

The article provides a thorough comparison of DSPy and Conva.AI, evaluating their strengths to determine which platform delivers the most sophisticated and efficient AI assistant solution. It also shares my personal experiences with both tools across various use cases, offering insight into their underlying processes.

Inspired by LMSYS, benchmarking platform to evaluate and compare the prompt generation capabilities of Conva.AI against others.

The article provides a complete walkthrough of Agent Arena, the story behind building it, comparing Conva.AI against other competitors like DSPy, Claude (Anthropic Workbench), Human built prompts and the totally unbiased approach it follows.