VIVEK NARAYANA

FastAPI gateway in front of vLLM with an OpenAI-compatible chat API and streaming. Gateway-side priority batching (short windows, premium first; each call is still a worker request—vLLM does GPU continuous batching internally), KV-cache-aware routing with optional Redis affinity, eviction and reroute under pressure, conversation-scoped exact and semantic prompt caching, in-memory tenant model/context limits, in-process SLO-style admission (429 on free tier when guards trip), Prometheus /metrics, and Kubernetes manifests with CPU-based HPA (metrics-server in cluster).

Built a modular Python framework for automating red teaming and security evaluation of LLM agents using Google Gemini. Features stateful multi-turn tool interactions via LangGraph, dynamic prompt injection attacks with meta-LLM payload generation, and automated evaluation using LLM-based judges to assess agent vulnerabilities and logical consistency.

Built an AI system that detects fires in real-time from surveillance footage using Roboflow, analyzes severity with Gemini, and autonomously contacts emergency services through a custom AI agent. Includes a FastAPI backend, real-time annotation, and alert logic.

Developed a sophisticated AI agentic loop that executes code in isolated Docker containers with real-time tool calling capabilities. Features MCP (Model Context Protocol) integration for GitHub API access and streaming response handling with comprehensive logging and error management.

Developed an AI-powered behavioral interview simulation feature using Groq, integrating facial expression and body language analysis with HumeAI. The app tracks performance using a React dashboard and stores results in Supabase.

Developed an AI-powered social media platform enabling users to engage in political discussions. Worked with a 4-person team to integrate PineconeDB for vector embeddings and used OpenAI/You.com API for summarizing legislation.

Developed a fashion website using OpenCV, Flask, and fuzzy logic to dynamically curate outfit schedules. Hosted on AWS with EC2, S3, and RDS using MySQL.

Applied OpenCV and YOLO to track ball possessions, player speeds, and distances in soccer clips. Utilized K-Means clustering to differentiate team colors and generate detailed player metrics.

Developed an algorithmic trading system that executes buy and sell orders for ETH, BTC, and LTC based on a custom strategy. The system automates trade execution and real-time order placement, sending instant notifications via a Discord bot.

Built a comprehensive calculus calculator capable of solving problems from Calculus I and II, including derivatives, integrals, limits, and series expansions. Developed during a hackathon (2/24 - 2/26), the project features a Flask backend for computation and a frontend using HTML, CSS, and JavaScript for an interactive user experience.