HIMAS: Healthcare Intelligence Multi-Agent System

Built a production-grade federated learning MLOps platform enabling privacy-preserving ICU mortality prediction (ROC-AUC 0.85) across 3 simulated hospitals using Flower, TensorFlow, and Airflow on GCP. The system trains on ~60K ICU stays from MIMIC-IV without centralizing patient data - each hospital's data stays local while a Flower SuperLink aggregates model weights via FedAvg over 15 rounds. End-to-end pipelines handle data ingestion, 35-feature engineering, Keras Tuner hyperparameter optimization, bias detection with Fairlearn, and automated CI/CD deployment to Cloud Run with rollback capabilities.

A 6-agent clinical AI system built with Google ADK and Gemini 2.5 provides real-time decision support - including mortality risk scoring, treatment optimization, resource allocation, and cross-hospital transfer coordination. All inter-hospital queries are privacy-preserved with K-anonymity (k=5) and differential privacy (ε=0.1), ensuring full HIPAA compliance with complete audit trails in BigQuery. [Won 1st Place at the Google Cambridge MLOps Hackathon.]

Use Cases: HIMAS enables hospitals to collaboratively improve ICU mortality prediction without sharing raw patient data - critical for multi-site health systems, clinical research networks, and any healthcare setting where data privacy regulations prevent centralized model training. Clinicians get real-time risk stratification, transfer recommendations, and network-wide case consultation while maintaining strict HIPAA compliance.

ClassifyAI: A Multi-Agent LLM System for Explainable Classification Pipelines

Architected a production-grade multi-agent LLM system with 8 specialized AI agents orchestrating end-to-end ML classification pipelines via LangGraph, FastAPI, and a Next.js frontend with real-time WebSocket tracking. The Double-Layer Architecture combines hardcoded statistical analysis (Layer 1) with LLM-generated custom code executed in secure Docker sandboxes (Layer 2), training 9+ algorithms (XGBoost, LightGBM, CatBoost, scikit-learn) in parallel - achieving 83.6% on clinical datasets and generating 5 deliverables per run: cleaned data, trained model, Jupyter notebook, technical report, and visualizations.

Human-in-the-loop approval gates at critical decision points (imputation strategy, feature selection, model choice) ensure full user control, while an interactive Project Manager agent provides real-time educational explanations of every transformation. All LLM-generated code is validated and sandboxed with no network access, complete audit trails, and resource-limited Docker execution.

Use Cases: ClassifyAI is built for data scientists and ML engineers who need rapid, explainable classification prototyping - from healthcare mortality prediction and fraud detection to customer churn analysis. Particularly valuable in regulated industries (healthcare, finance) where reproducibility, audit trails, and automated documentation are mandatory.

Jackie x AI Voice Assistant

Built a production-ready AI voice assistant using FastAPI and WebSocket for real-time communication, leveraging Groq's ultra-fast Llama-3.3-70B model for intelligent responses. Implemented comprehensive speech-to-text using Groq Whisper with local fallbacks, and high-quality text-to-speech synthesis using Microsoft Edge TTS for natural voice interactions.

Deployed on Modal.com serverless infrastructure with encrypted API key management and DuckDuckGo web search integration for current information retrieval. The system achieves 3-7 second end-to-end response times with automatic scaling and maintains secure HTTPS/WSS communication protocols.

Use Cases: Professional portfolio assistance for recruiters and clients, real-time voice-enabled Q&A for accessibility, automated customer support systems, and intelligent virtual receptionists for businesses across healthcare, education, and technology sectors.

NewsBuddy AI: Article-Grounded News Companion

Built a full-stack AI news companion that turns articles into interactive conversations. The system combines a Next.js 14 frontend with a FastAPI backend, opening a per-article WebSocket session that streams an initial summary and answers follow-up questions using article-grounded retrieval instead of drifting into generic chatbot responses.

Implemented retrieval-augmented generation with FAISS-backed sentence-transformer embeddings over article chunks, Groq-powered LLM responses, optional Tavily web snippets, and news ingestion from RSS/provider APIs. The companion supports typed questions, browser speech-to-text, spoken replies through Groq TTS with Edge TTS fallback, and latency instrumentation for first-assistant-response timing.

Use Cases: NewsBuddy helps readers understand complex stories faster by generating concise briefings, preserving source context, supporting voice-enabled Q&A, and linking answers back to real article sources. It is useful for daily news discovery, research workflows, accessibility-focused reading, and source-grounded media assistants.

PaperPal AI - will get you direct answers from the pdf documents without any addition information

Developed a Streamlit-based application utilizing Google's Gemini-PRO model within a Retrieval-Augmented Generation (RAG) framework to extract, analyze, and generate contextually accurate and specific insights from PDF documents.

Integrated features for instant insight generation from PDF uploads and implemented a secure API key input functionality

Industry Use Cases: PaperPal AI excels in scanning legal documents, providing accurate and direct answers to legal cases. It enhances efficiency by quickly processing extensive texts, allowing legal professionals to focus on critical analytical tasks.

Real-time Face Mask Detector

Developed a highly accurate deep learning model for real-time face mask detection, achieving an accuracy rate of 84% using an extensive and diverse image dataset.

Implemented the model in a real-world application by integrating it with webcam systems, demonstrating its effectiveness in monitoring and ensuring public safety.

Conducted rigorous testing and validation in various environments, showcasing the model's reliability and potential for deployment in public spaces and safety-critical scenarios.

Use Cases: Can be installed on dashcams to monitor drivers, enhancing safety and preventing major accidents by ensuring drivers are alert and compliant with safety protocols.

Driver Drowsiness Detection System

Engineered a driver drowsiness detection system using Machine Learning, OpenCV, and Convolutional Neural Networks (CNN), achieving an accuracy of 87% and integrating an alarm for immediate driver alert.

Improved system reliability by reducing false positives by 10%, enhancing the accuracy and trustworthiness of the drowsiness alerts, and ensuring safer driving conditions.

Conducted extensive testing and optimization, ensuring the system's robustness and effectiveness in various real-world driving environments and lighting conditions.

Use Cases: This system reduces false positives by 10%, significantly enhancing driver safety and preventing major accidents on the roads caused by driver drowsiness.

VideoLLM: Multi-Modal Video Ad Classifier

Created a multi-modal classifier that integrates text (DistilBERT) and visual data (ResNet3D) to classify video advertisements based on 21 yes/no questions. Achieved an F1 score of 0.8088 and an accuracy of 0.8184 on test data, excelling in identifying brand mentions, calls to action, and emotional impact.

Utilized EasyOCR for on-screen text extraction, Whisper for speech-to-text, and DistilBART for summarization, enabling robust analysis of real-world ads that include spoken dialogue, branding visuals, and complex marketing language.

Use Cases: Automating analysis of large-scale video ad campaigns for brand compliance checks, emotional impact assessments, and content strategy insights, ultimately helping marketers optimize ad content and improve targeting strategies.

Predictive Modeling for Property Interaction

Developed a predictive model to forecast user interactions with properties based on various features such as property details, interaction history, and photo counts. The model uses classical machine learning techniques to predict interactions within 3 and 7 days, offering insights into user engagement with real estate listings.

The project involved thorough data preprocessing, including outlier removal, one-hot encoding, and feature scaling with MinMaxScaler. The dataset was composed of property data, interaction logs, and photo data, and multiple models were trained and evaluated to optimize prediction accuracy.

Use Cases: The predictive model helps real estate platforms and property agents by forecasting which properties are likely to receive more interactions, assisting in targeted marketing and property management strategies.

Company Similarity Finder Web Application

Developed an innovative web application to find and explore companies similar to a given company, based on their descriptions. The app utilizes advanced NLP techniques, particularly Sentence-BERT (SBERT), to generate embeddings of company descriptions and uses FAISS for efficient similarity searches on large datasets.

The project efficiently handles large datasets (55,000+ companies) by caching precomputed embeddings and utilizing FAISS for fast nearest neighbor searches. The app is built with Streamlit, providing an intuitive user interface for exploring company similarities and visualizing results.

Use Cases: This tool enables users, such as investors or market researchers, to find companies with similar business models, products, and services, streamlining market analysis and helping with strategic decision-making.