Aayush Joshi

About

Aspiring Machine Learning Engineer with hands-on experience in developing AI-driven solutions, specializing in Python, PyTorch, and Transformer models. Adept at building and fine-tuning LLMs and implementing computer vision systems. Strong problem-solving abilities, automation expertise, and a passion for creating impactful machine learning projects. Proven track record in leveraging AI for search summarization, finance automation, and emotion detection.

I am open to collaborating on projects, and I'm always looking for new opportunities to learn and grow.

Experience

Software Engineer

ThinkAct AI

Remote

2025/06 - Present

Architected and deployed advanced RAG applications with multi-tenant architecture and highly configurable, decoupled components using YAML-based service configurations, scalable vector database integration, robust task management, and optimized retrieval methods. Designed and implemented AI-driven platforms and specialized agents leveraging LangChain, LangGraph, and Model Context Protocol (MCP), including task-specific MCP servers and client integrations. Led a small development team, conducted peer code reviews, and contributed to backend systems using SQL/NoSQL databases and FastAPI-based RESTful APIs.

Software Engineer Intern

Dolphant Group LLC

Remote

2025/04 - 2025/06

Fine-tuned open-source large language models (LLMs) on domain-specific datasets using QLoRA and Reinforcement Learning from Human Feedback (RLHF) on AWS. Developed scalable RESTful APIs with FastAPI to integrate LLMs and other backend services.

Machine Learning Intern

The Social Purpose Trust

Remote

2025/02 - 2025/04

Developed AI-driven models for climate solutions, including predictive analytics for environmental monitoring and impact assessment. Collected, preprocessed, and augmented climate-related datasets to improve model performance. Automated climate data analysis and reporting, integrating solutions into the Surya Sangam Web Portal to support sustainability initiatives.

Projects

Interviewer AI Agent

Developed an AI agent that conducts interviews with users based on there resume and job description. The agent uses a LLM to generate interview questions, both coding and experience based, and evaluates the user's responses.

Python

Langgraph

Langchain

FastAPI

PostgresSQL

Llama-3.3-70b-specdec

Snake Game

Snake Game with AI combines the classic Snake game with an AI agent trained using reinforcement learning. It features both manual and AI-controlled gameplay, customizable settings, score tracking, and real-time plotting of training progress. The AI uses a Q-network built with PyTorch to navigate the game.

Python

PyTorch

Pygame

NumPy

Matplotlib

Personal Portfolio

Developed a personal portfolio website using Next.js

Next.js

Tailwind CSS

Surch AI

Developed a web-based search summarizer using Google Custom Search API, Beautiful Soup, and Langchain for conversational retrieval. Implemented ChromaDB for vector storage and added a history feature with SQLite to track and revisit user queries.

Python

Langchain

Streamlit

Langchain-Huggingface

SQLite

ChromaDB

Mistral-7B-v0.3

Finance LLM

Developed Python scripts for automated data collection using Google Custom Search API and built an LLM dataset parser to generate a QA dataset. Fine-tuned an open-source Hugging Face model with the custom dataset to optimize performance. Utilized libraries like Beautiful Soup, PyPDF2, and Python-dotenv for auxiliary tasks.

Python

Transformers

TRL

Bitsandbytes

Accelerate

Google Colab

Emailer

Emailer Project automates sending emails to recruiters using the Gmail API. It requires setting up Gmail API credentials and a list of recruiter emails in an Excel file. Once configured, the project sends emails with an attached resume.

Python

Pandas

Openpyxl

Google-auth

Gmail API

Chat with pdf

Built a Streamlit web application that allows users to upload a PDF file and interact with its content using a chat interface. It leverages LangChain for document retrieval and processing, and HuggingFace for embeddings.

Python

Streamlit

Langchain

Huggingface

Loan Approval

Performed basic preprocessing and EDA on a loan dataset.

Python

Pandas

Scikit-learn

Matplotlib

Seaborn

Face Emotion Detection

Integrated YOLO for face detection and fine-tuned ResNet18 for classifying seven emotions: Angry, Disgust, Fear, Happy, Neutral, Sad, and Surprise. Used pre-trained models for quick inference and easy setup on Linux systems.

Python

PyTorch

YOLO

ResNet

OpenCV

Linux

Epsilon

Epsilon is a 2D physics engine for reinforcement learning, simulating interactions with shapes like rectangles and circles. It features collision detection, response, and basic force application. Currently in development, it aims to improve collision handling and integrate more fully with RL, though issues with circle and rod collisions remain.

Python

Pygame

© 2024 Aayush Joshi