SignBridge
AI-Powered ASL Accessibility & Learning Platform
A full-stack accessibility and learning platform for American Sign Language, built around three integrated modules — real-time computer-vision fingerspelling recognition, everyday sign communication, and a gamified learning curriculum — designed around how different learners and signers actually use ASL.
Explore the three modules
Webcam-based ASL fingerspelling recognition (A–Z) using MediaPipe hand-landmark detection and a scikit-learn MLP classifier trained on real image data — ~86% accuracy on a held-out test set (small sample; not yet validated across real-world webcam conditions).
Types English text and plays back verified sign video for supported vocabulary, with fingerspelling as a fallback for everything else — a supported-vocabulary communication aid, not a full grammatical English-to-ASL translator.
Gamified curriculum — XP, levels, streaks, hearts, and achievements — covering the full A–Z alphabet plus a verified vocabulary library, with real signer videos and source attribution, comprehension quizzes, and camera practice validated against the recognition model. Actively expanding with per-sign mastery tracking, quests, and contextual practice scenarios.
The recognition model is objectively validated on A–Z fingerspelling only; vocabulary-sign practice in Learning is self-guided and self-assessed, not model-graded.
OrderFlow
Built six independently deployable Spring Boot microservices coordinating asynchronous order, inventory, payment, and notification workflows through Kafka.
Designed reliable event processing with transactional outbox, idempotent consumers, retries, dead-letter handling, and compensating transactions.
Implemented concurrency-safe inventory reservation and automated integration testing with JUnit, Testcontainers, Docker Compose, and GitHub Actions.
AI-Powered Job Application Tracker
Full-stack application for tracking job applications, interview stages, deadlines, and history — React frontend, FastAPI/PostgreSQL backend.
Authentication-protected workflows with filtering and search over application records, exposed via REST APIs.
AI-assisted analysis of job descriptions to surface relevant skills; containerized with Docker and deployed on AWS.
NeuralMed
Flask-based application applying multiple disease-screening models across structured clinical data and imaging.
Kidney CT classification via a fine-tuned VGG-16 model, alongside classical ML (scikit-learn) and deep learning (TensorFlow) pipelines for other conditions.
Integrated Gemini-generated clinical summaries alongside model outputs.
Ferry Loading Optimizer
Python algorithm optimizing vehicle loading onto a ferry under size and weight constraints.
Achieved up to a 30% reduction in unused capacity in simulated loading scenarios.
Improved runtime by 40% versus the initial implementation through data-structure optimization.
AI-Powered ECG Signal Processing & Arrhythmia Detection
Streamlit application for ECG signal processing, PQRST peak detection, and CNN-based arrhythmia risk classification (Normal / Moderate Risk / High Risk).
Trained and evaluated on the MIT-BIH Arrhythmia Database (PhysioNet) using WFDB for signal loading; achieved 98.8% classification accuracy.
Containerized with Docker and tested via GitHub Actions CI.
Ethical Challenges of AI in Hiring — Bias Audit Study
Collaborative project with Lekha Dakshinamurthy and Sathyadharini Srinivasan, auditing a hiring model trained on a 2,000-applicant synthetic dataset.
Found that surface-level fairness metrics passed regulatory thresholds while 24.8% of hiring decisions changed when only gender was flipped — revealed using SHAP, LIME, and counterfactual analysis.
Proposed the IDEAL framework (Inclusive Data, Debiased Models, Explainability by Default, Auditing & Monitoring, Legal Compliance) for ethical AI recruitment.
Plant Detection using YOLOv8
Flask web application identifying medicinal/herbal plant species (Neem, Oleander, Ficus Religiosa, Calotropis) from uploaded images using a trained YOLOv8 model.
Built on a public Roboflow plant-detection dataset (CC BY 4.0).
Returns the identified species along with basic traditional-use information.