Madhumitha
Chandrasekaran

Developed a deep learning pipeline to classify MRI brain scans into four Alzheimer's disease stages, followed by LSTM-based time-series forecasting to predict disease progression and cognitive decline patterns. Built a clinician-facing application prototype that surfaced classification results, progression forecasts, patient history, and diagnostic insights through an interactive interface.

• Four-class MRI classification (Non-Demented, Very Mild, Mild, Moderate Dementia) using a CNN-LSTM hybrid achieving 87% accuracy.
• Correlated CNN-extracted features with longitudinal OASIS dataset data to model disease evolution over multiple time points.
• Prepared MRI feature vectors into patient-level time-series sequences capturing temporal patterns including cortical thinning and ventricular dilation.
• Implemented LSTM networks to learn long-term sequential dependencies and avoid vanishing gradient problems.
• Correlated MRI features with cognitive scores (MMSE, ACE) and clinical ratings to build prediction patterns between brain atrophy and cognitive impairment stages.
• Built a forecasting model to predict future atrophy patterns and cognitive decline trajectories, supporting early intervention and personalized treatment planning.
• Developed the clinician-facing app: MRI upload and auto-classification (AD / CN / EMCI / LMCI), patient history tracking and confidence scores.
• Enabled side-by-side comparison of MRI-based classification with longitudinal assessments for comprehensive diagnosis support.
• Published in IEEE Xplore at the 2025 ICCDS.

Clinicians lack integrated tooling that links MRI-based staging with longitudinal cognitive forecasts. Existing models stop at classification and leave progression and intervention planning to manual interpretation.

Part of a research team. My contributions focused on time-series analysis, LSTM modeling and building the clinician-facing application interface. I correlated CNN-extracted features with the longitudinal OASIS dataset, prepared MRI feature vectors into patient-level time-series sequences, and implemented LSTM networks to learn long-term sequential dependencies. I also correlated MRI features with cognitive scores (MMSE, ACE) and clinical ratings to build prediction patterns, and developed the forecasting model to predict future atrophy and cognitive decline trajectories.

Classifies MRI scans into 4 Alzheimer's progression stages using a CNN-LSTM hybrid. Captures spatial patterns (CNN) and longitudinal progression signals across slices (LSTM). Performs image preprocessing and augmentation: normalization, resizing, dataset balancing. Provides a visual analytics interface for clinicians highlighting affected brain regions. Achieved 87% classification accuracy across dementia progression stages.

I developed the clinician-facing application with MRI upload and auto-classification (AD / CN / EMCI / LMCI), patient history tracking and confidence scores. It enables side-by-side comparison of MRI-based classification with longitudinal assessments for comprehensive diagnosis support.

Delivered a working prototype and an IEEE publication validating the pipeline. The interface lets clinicians move from a single scan to a personalized progression view in one workflow.

Moodify is a real-time emotion-driven music recommendation system that uses facial expression recognition to detect the user's current mood and automatically generate a Spotify playlist tailored to it.

• Integrated the Haar Cascade Classifier to detect faces from live webcam video frames in real time.
• Built a CNN-based emotion classification model trained on FER2013 (35,887 grayscale images) to classify 7 emotions: Angry, Disgust, Fear, Happy, Neutral, Sad and Surprise.
• Implemented frame-level emotion tracking across 50 frames to determine the dominant emotion per session.
• Handled image normalization, grayscale conversion and dimensional formatting for model inference.
• Ensured the detected emotion label was correctly saved and passed downstream to the playlist mapping pipeline.
• PCA reduced Spotify audio feature vectors to 6 components for efficient clustering.
• K-Means grouped Spotify tracks into 7 mood-based clusters using audio features.
• Detected emotion mapped to a mood cluster and a 26-song playlist was auto-created via the Spotify API.

Part of a team. My contribution focused on the emotion detection and classification model pipeline. I integrated the Haar Cascade Classifier to detect faces from live webcam video frames in real time, built the CNN-based emotion classification model trained on FER2013 (35,887 grayscale images) to classify 7 emotions, implemented frame-level emotion tracking across 50 frames to determine the dominant emotion per session, handled image normalization and grayscale conversion for model inference, and ensured the detected emotion label was correctly saved and passed downstream to the playlist mapping pipeline.

1) Haar Cascade Classifier detects faces from live webcam frames. 2) CNN classifies emotion into 7 states with a confidence score. 3) PCA reduces audio feature vectors to 6 components for efficient clustering. 4) K-Means groups Spotify tracks into 7 mood-based clusters using audio features. 5) Detected emotion maps to mood cluster and a 26-song playlist is auto-created via Spotify API.

Happy → Upbeat. Sad → Calming. Angry → High-energy. Disgust → Soothing. Surprise → Dynamic. Fear → Ambient. Neutral → Balanced.

Delivered an end-to-end application integrating real-time emotion recognition, music recommendation, and Spotify playlist generation. The system tracks emotion across frames, maps the dominant mood to a curated playlist, and creates it automatically, giving users music that matches how they feel.

AI-driven chess system capable of detecting board states, computing optimal moves and executing them physically via a robotic arm in real time. Funded by SSN Research Centre.

• Built the computer vision pipeline that translates the physical board state into a digital representation using OpenCV: chessboard grid detection, square segmentation and piece recognition in real time.
• Converted detected board configurations into FEN notation for integration with the Stockfish chess engine to compute optimal moves.
• Contributed to designing the pipeline that maps AI-generated moves to board coordinates, enabling the system to translate strategic decisions into physical piece navigation.
• Demonstrated how computer vision, AI decision systems and physical game environments can be integrated to create interactive intelligent systems.

• OpenCV detects chessboard grid, identifies piece positions and converts board state to FEN notation.
• Stockfish engine computes optimal moves from the detected board configuration.
• State-to-action pipeline maps digital moves to physical board coordinates for robotic navigation.
• Real-time board recognition updates game states dynamically as moves are played.
• Demonstrates full integration of computer vision, AI decision systems and physical automation.

An end-to-end AI-powered analytics tool that converts plain English business questions into live SQL queries and actionable insights. Built out of a personal interest in applying emerging AI technologies to real-world data workflows.

• Accepts a natural language business question via a branded dashboard.
• Claude AI API generates optimized PostgreSQL from the question automatically.
• Query executes live against Supabase PostgreSQL (Olist dataset, 100K+ orders, 9 relational tables).
• Returns structured analysis: plain-English summary, data quality audit and visualization recommendation.
• n8n orchestrates the full workflow end-to-end with no manual intervention.

Conducted a rigorous A/B test comparing a minimal system prompt against a schema-aware engineered prompt across 20 diverse business questions, measuring SQL accuracy, query correctness and output reliability. The engineered prompt achieved significantly higher accuracy, providing evidence-based justification for the final system design.

• End-to-end NL-to-SQL pipeline with live database execution and structured output.
• Schema-aware prompt engineering validated through experimental A/B design.
• Demonstrates proficiency in AI tool integration, automated pipeline development and translating technical outputs into business-ready insights.
• Custom-built branded dashboard for non-technical business users.

Built to answer one question: what am I actually paying for, and is it still worth it? SubSight scans Gmail for subscription, billing and trial emails, uses Claude AI to extract structured billing data, verifies service status via Exa web search, and delivers a formatted weekly report autonomously, with no manual intervention required.

• Scans Gmail for subscription, billing, receipt and trial emails from the last 30 days.
• Uses Claude AI to extract service name, price, billing cycle and trial dates from raw email content.
• Detects price changes between billing cycles and alerts you once in the next report.
• Flags free trials ending within 7 days before your card gets charged.
• Verifies via Exa web search whether active services are still operating.
• Sends a formatted weekly report every Sunday at 8am with active spend, alerts and cancelled services.
• Runs autonomously via a macOS LaunchAgent, no terminal required after setup.

• End-to-end automated pipeline: Gmail → Claude extraction → Exa verification → weekly email digest.
• Handles edge cases: re-subscriptions after cancellation, price change deduplication, trial expiry warnings.
• REST API (FastAPI) with trigger endpoints for on-demand scan, verify, report and full pipeline runs.
• Demonstrates real-world application of LLM extraction, workflow automation and API orchestration.