My Featured Projects

AI chatbot web app with a clean web interface and a secure backend. Users can sign in, create projects, upload files, and chat in real-time with fast LLM models provided by OpenRouter, while provider API keys stay safely on the server. It’s built for reliability and growth with streaming responses, rate limiting, usage tracking, and an event-driven architecture for scalability.

• Implemented JWT and OAuth2-based authentication for protected user sessions.
• Added file upload and retrieval workflows to ground model responses to the file context.

Call-to-CRM AI Assistant reads a sales call transcript, suggests CRM updates, drafts follow-up tasks and an email, and waits for human approval before applying any changes. It is designed to make post-call CRM work faster without allowing the model to write directly to the CRM.

• Accepts a pasted transcript, meeting notes, or a plain text upload.
• Loads the current opportunity from a seeded demo CRM.

A Financial Analysis tool that combines traditional metrics visualization with AI-powered insights. This application helps financial analysts and business owners track key revenue metrics and generate automated explanations using Large Language Models (LLMs).

• Interactive Dashboard: Built with Streamlit for real-time interaction.
• Key Metrics Visualization: Track important financial metrics

Applied machine learning workflow for anticipating maintenance needs with experiment tracking and a user-friendly web UI.

• Trained and evaluated predictive models for maintenance forecasting.
• Used MLflow for experiment tracking and model versioning.

A Multi-Agent System built with CrewAI that helps users tailor their resumes and prepare for interviews based on specific job postings. Consists of 4 AI agents with well-defined roles, tasks and tools.

• Tech Job Researcher: Analyzes the job posting to extract key requirements.
• Personal Profiler: Compiles a comprehensive profile from user's GitHub and personal write-up.

A semantic book recommendation system designed to find books based on natural language queries, categories, and emotional tones. Unlike traditional keyword search, this system understands the semantic meaning of your query.

• Semantic Search: Finds books based on the content of their descriptions rather than just title matches.
• Emotion Analysis: Filters books by emotional tone (Happy, Surprising, Angry, Suspenseful, Sad).

This project provides a pipeline for fine-tuning the Phi-3-mini-4k-instruct model for Information Extraction tasks. By leveraging the Unsloth framework and LoRA (Low-Rank Adaptation), we enable efficient fine-tuning on consumer-grade GPUs, transforming unstructured text into structured JSON data.

• Efficient Fine-Tuning: Uses Unsloth's optimized kernels for 2x faster training and 60% less memory usage.
• Configurable Pipeline: All training parameters (learning rate, epochs, LoRA rank) are externalized in config.yaml.

An implementation of the Transformer architecture (specifically a Decoder-only GPT-style model) built from scratch in PyTorch. Trained on the Shakespeare dataset from Kaggle to generate Shakespearean-like text.

• Architecture: Decoder-only Transformer (suited for text generation).
• Components: Multi-Head Attention, Sinusoidal Positional Encoding, Feed-Forward Networks, Layer Normalization.

Python-based Model Context Protocol server that lets LLMs perform CRUD operations on PostgreSQL data through structured tools.

• Exposed database CRUD operations through MCP-compatible tools.
• Focused on reliable query execution so LLM-driven workflows could interact with live relational data.

A command-line tool that automatically retrieves your unread Gmail messages and generates concise summaries using a local AI model. The tool processes emails intelligently, handles long content gracefully, and marks emails as read after summarization.

• Built a CLI-based workflow, UI planned for the next version.
• No external API costs - runs entirely offline after initial setup