Khagendra Khatri

A cloud-based synthetic cortex that enables memory-grounded, identity-preserving AI social media automation through transparent autonomous agents.

Built and launched core backend infrastructure using FastAPI, PostgreSQL, and gRPC, scaling to 1,000+ active users. Designed real-time data aggregation pipelines and optimized caching layers with Redis.

Founded a software company delivering 4 AI-powered iOS apps with scalable FastAPI backends. Engineered a custom NGINX-based load balancer on Render to distribute traffic, reducing costs by 95%.

A digital memorial platform that lets users create AI-powered tributes for loved ones. Built during HackRU, it combines modern web tech and AI to enable meaningful interactions and conversations with those who have passed.

Browser extension for detecting AI deepfakes, explicit content, and misinformation.

AI-powered math visualizer and intuition-building learning tool that transforms concepts into immersive animations using Manim.

Website/app focused on therapeutic engagement and mental well-being.

Assistive computer-vision platform for visually impaired navigation.

Digital AI interaction platform enabling conversations with virtual representations of loved ones.

Built backend for an AI-powered interview agent using FastAPI. Orchestrated multi-step interview workflows using LangGraph and implemented Redis-based session persistence.

Trained an action-based Recurrent Neural Network for ASL translation with Pascal tree style nodes. Achieved 19% accuracy on 100 different words.

Led development of AI solutions using OpenAI Assistant API and Google Gemini. Managed end-to-end data pipelines, improving model performance by 25%.

Created a modified GRU from scratch using numpy, achieving 10% performance improvement. Invented Recurrent Symbolic Networks to tackle vanishing gradients.

Developed backend services for cashback/rewards platform using Python and PostgreSQL. Optimized ETL workflows, reducing manual processing overhead by 70%.

Trained a Neural Network with 550k parameters using TensorFlow/Keras, achieving 95% accuracy in detecting assignment plagiarism. Published in 28th Annual CCS Conference.