Current Projects
MindMate – Agentic AI Platform for Mental Health Assessment
Mental health assessments are often slow, unstructured, and lack personalized follow-ups. To address this, I designed MindMate, a platform that streamlines preliminary diagnosis while ensuring reliability and structured outcomes.
- •Developed a multi-agent pipeline powered by LLMs for assessment, diagnosis, and specialist matching
- •Integrated the standardized SCID tool to build the assessment pipeline, and applied grounded reasoning with LLMs to infer conditions using DSM-5 evaluation criteria
- •Delivered a Dockerized FastAPI prototype, enabling structured AI-assisted evaluations that reduce manual effort and generate actionable reports for clinicians
Tech Stack: Python, LangChain, FastAPI, Docker
BookMate – RAG-Based Intelligent Document Explorer
Professionals often struggle to query multi-format documents while also needing access to real-time external knowledge. I built BookMate, a system that unifies document understanding with live search in a single workflow.
- •Implemented a ReAct (Reasoning + Acting) agent to dynamically choose between document retrieval (via ChromaDB) or live web search (via ScraperAPI)
- •Designed a RAG pipeline using sentence-transformers for embedding-based retrieval
- •Delivered an interactive assistant capable of handling PDFs, DOCX, and TXT files, providing seamless querying across internal documents and external knowledge sources to improve accuracy and reduce search effort
Tech Stack: Python, LangGraph, ChromaDB, Groq, ScraperAPI, Sentence-Transformers
ML-Based Dysarthria Detection System
Early detection of speech disorders like dysarthria can significantly improve treatment outcomes, but manual screening is resource-intensive. I developed an ML-based system to classify speech as normal or abnormal, supporting clinical assessments.
- •Extracted speech embeddings with Wav2Vec2 and engineered acoustic features using Librosa
- •Trained and compared multiple models (Scikit-learn, TensorFlow) for robustness and reliability
- •Designed real-time inference capabilities for integration into clinical workflows, achieving 94% classification accuracy and demonstrating the system's potential as a reliable AI assistant for clinicians
Tech Stack: Python, Hugging Face Transformers, Librosa, TensorFlow, Scikit-learn
Future Projects
Personalized Career Mentor (Planned)
Many students and early professionals struggle with career decisions because guidance is often generic and disconnected from their unique skills and aspirations. I set out to design an AI system that could provide personalized, evolving mentorship.
- •Conceptualized an AI-powered career mentor that analyzes a user's skills, interests, and goals to generate tailored learning paths, job recommendations, and mentorship matches
- •Designed the system around LLMs + recommendation engines + skills ontologies, enabling adaptive guidance that improves as the user progresses
- •Planned to integrate real-time career data sources (job boards, course catalogs, mentorship networks) to keep advice practical and relevant
Planned Tech Stack: Python, LangChain, FastAPI, Hugging Face Transformers, Vector Databases
ClassMate – AI Tutor That Reads Student Expressions
Online education often misses the personal feedback loop of a real classroom — teachers can't always tell if a student is confused, bored, or engaged. I wanted to design a system that brings that 'human awareness' into digital learning.
- •Built an AI-powered tutor that analyzes facial expressions and gaze to detect emotions like confusion, boredom, and engagement
- •Integrated FER+ and AffectNet models for emotion recognition and attention-tracking algorithms for gaze detection
- •Designed adaptive feedback loops where the system flags struggling learners and adjusts the lesson flow in real time
Planned Tech Stack: Python, OpenCV, PyTorch, FER+, AffectNet
Interested in collaborating or learning more about my work? Feel free to reach out or explore more of my projects and insights.