Varshini Venkataraman

About Me

I build systems that hold up under real load, not just demos. I have an MS in Computer Science from UMass Amherst (graduated May 2026) and I'm looking for full-time Software Engineering roles across the US.

At Lillup, I fixed a Next.js app's security gaps and built a CI pipeline that gates every pull request on ~90% test coverage. On my own time, I've built and load-tested platforms with Spring Boot, Next.js, and Kafka — one handling 1,300+ requests per second with zero errors, another passing every Gatling load test I threw at it. I test for failure before it happens, and that habit carries over into how I build with AI too.

That's the same instinct behind Kira and Penny — Kira is a multi-agent system, built with a teammate, that turns a single WhatsApp message into a published YouTube Short using Google ADK, Gemini, and tool-calling orchestration; Penny is an agentic finance app built on Claude's tool-use API with agent-to-agent delegation and a working MCP server (more below). I've also fine-tuned LLaMA-3.2-3B with LoRA during an IBM/UMass externship program.

I like problems where correctness actually matters — where "it works on my machine" isn't good enough, and someone's data or money is on the line. If that's the kind of team you're building, I'd love to talk.

Education

Technical Skills

Experience

Projects

Kira — An Autonomous AI Agent That Runs Your YouTube Channel

• Collaborated in a team of two to build an autonomous multi-agent AI system using Google ADK, Gemini, and LLM orchestration that transforms a single WhatsApp request into a publish-ready YouTube Short, using tool calling for trend research, script generation, AI image/video generation, TTS, music, Whisper-synchronized captions, and automated publishing.
• Engineered the backend and production infrastructure with Python, FastAPI, Twilio, PostgreSQL, FFmpeg, and Google Cloud Run, implementing persistent agent memory, asynchronous processing, reliable agent handoffs, proactive WhatsApp updates, rate limiting, privacy controls, and GitHub Actions CI/CD.

Google ADKGeminiMulti-Agent SystemsPythonFastAPITwilioFFmpegGoogle Cloud Run

Live Demo → GitHub

Penny — AI-Powered Personal Finance Agent

• Built a multi-tool agentic chat assistant on Claude's tool-use API (SQL generation, full-text search, dynamic charting, code execution, web search), correctly handling parallel and multi-step tool calls with accurate conversation-state reconstruction across turns.
• Implemented agent-to-agent delegation: the primary chat agent dispatches ambiguous merchant-categorization requests to an independent sub-agent with its own scoped system prompt and fresh context window, persisting results back to the shared ledger.

PythonStreamlitDuckDBClaude APITool UseMCPSub-AgentsAgent SkillsPlotly

GitHub

VaultFlow — Banking Platform

• Architected a production-grade banking platform with 5 decoupled Spring Boot 4 microservices and a Next.js 14 frontend, featuring event-driven fraud detection, AI-powered transaction categorization, and a full observability stack (Prometheus, Grafana, Loki, Zipkin) deployed at $0/month.
• Engineered JWT auth with refresh token rotation and idempotency-key deduplication through Spring Cloud Gateway, with Kafka-driven fraud and notification pipelines validated via Gatling load tests — 100% success rate, 89ms mean latency, and sub-13ms read response times across 6 banking operations.

Spring Boot 4Next.js 14PostgreSQLKafkaResilience4jObservability Stack

GitHub

Live Demo

CampusTix — Event Ticketing Platform

• Built concurrent seat-booking with optimistic locking (@Version), validated zero duplicate bookings under 500 concurrent users via JUnit and CountDownLatch, and load-tested with k6 achieving P95 latency of 93ms at 1,306 req/s with zero 5xx errors across 260K requests.
• Implemented async booking via @Async and STOMP WebSocket push returning 202 in ~4ms median, with Redis rate limiting, Resilience4j circuit breaker isolating AI faults across 4 fault-injection scenarios, and Micrometer instrumentation deployed via multi-stage Docker on Railway.

Java 23Spring Boot 4PostgreSQLRedisWebSocketDocker

Live Demo → GitHub

Disk-based Database Engine

• Implemented Page and Row serialization, LRU buffer pool management with disk persistence via RandomAccessFile, and a BTreeLeafNode with point and range search — achieving 300–3000× speedup over full table scans on IMDB dataset queries.
• Assembled a pipelined query executor with CLI interface supporting Block Nested Loop Join across Movies, WorkedOn, and People relations, as part of a 4-person team.

JavaB+ TreeBuffer ManagementBNLJ

GitHub

Certifications

Publication

IC3IoT 2024

Weed Plant Detection in Agricultural Field Using Deep Learning Integrated with IoT

2024 5th International Conference on Communication, Computing and Internet of Things (IC3IoT)

Read on IEEE Xplore →

Get In Touch

Open to full-time Software Engineering roles. Let's build something great together.