Selected work in conversational AI, computer vision, semantic search, and ML for cybersecurity — from multilingual voice agents to deepfake detection, vibe-based discovery, and automated YARA rule generation.
A real-time voice agent that handles Hindi/Bengali ↔ English code-switching — mixing languages mid-sentence the way Indian users actually speak — instead of forcing a single-language mode. Users upload a PDF and query it by voice, with answers grounded only in that document.
An ensemble deepfake classifier (XceptionNet + EfficientNet) achieving 96.72% test accuracy, 91.96% F1-score, and 97.57% specificity on a balanced 8,538-image subset of FaceForensics++ (FF++). Deployed as a Dockerized FastAPI inference service on Hugging Face.
An AI-powered movie discovery platform enabling natural-language “vibe” search in place of keyword or title search, using Sentence-BERT embeddings with a hybrid retrieval and similarity-ranking pipeline. Full-stack architecture with Python backend, Next.js frontend, and PostgreSQL storage.
An AI-driven system for automated YARA rule generation, built during ML-for-cybersecurity research at Texas A&M under Prof. Marcus Botacin. Apriori extracts detection patterns from malware binaries and security telemetry; adaptive stream-learning models keep those patterns updating continuously instead of relying on static batch retraining.
Whether you have a specific project in mind, want to discuss machine learning architectures, or just want to network, my inbox is always open.