Developed a scalable EdTech platform supporting multilingual learning, transitioning the product from MVP to a production-ready application.

Built over 50+ RESTful APIs (Node.js/Express.js) and a React.js + Redux + Tailwind based frontend, deployed on Vercel & Render, successfully handling 5,000+ concurrent users with less than 120ms API latency and 99.98% uptime.

Optimized MongoDB with sharding & indexing, reducing query times by 40%, and integrated NGINX load balancing & asset optimization to cut Time to Interactive (TTI) by 35%.

Built a LangGraph-based Multilingual & Multi-Modal RAG Chatbot (document, text & audio) integrated with Qdrant & hallucination-mitigation techniques, enabling indigenous-style syntax and grammar learning across 15+ languages, while improving query resolution speed by 60% & lowering fallback rates to under 5%.

Designed CI/CD pipelines with Docker & GitHub Actions, reducing deployment time from 15 min to under 3 min and enabling zero-downtime releases.

Integrated Namecheap Email Services for 99% reliable email delivery, JWT + OAuth 2.0 for secure authentication, and Stripe for payments, achieving 100% PCI-DSS compliance.

Collaborated in an Agile team, contributing to code reviews, feature planning, and iterative development cycles.

Designed and implemented a multimodal deepfake detection pipeline combining EfficientNet-B0 for video and Wav2Vec2 for audio to capture joint audio-visual representations.

Improved fake/real classification accuracy by around 16% (from 78% to 94%) by leveraging cross-attention fusion for robust feature alignment across modalities.

Processed and analyzed 10,000+ video samples from the DFDC dataset, extracting synchronized face frames and audio spectrograms to enable large-scale training.

Developed PyTorch-based CNN & Transformer models, optimizing with Adam and learning-rate schedulers, achieving less than 120ms inference latency per video clip.

Contributed to a research publication on synthetic media detection, addressing real-world threats of misinformation through unified multi-modal deep learning.