Varshitha Gogineni

Alexa+ style voice add-on built for Amazon's Build, Ship, Shape hackathon — a self-hosted MCP server (spec 2025-11-25) on Amazon Bedrock AgentCore Runtime that talks you through appliance repairs step by step, hands-free, with every instruction served verbatim from the manufacturer's own manual so the model can never invent a repair step.

• Engineered code-enforced voice safety: a power-off safety gate that clears only on the user's own words, checked server-side — measured over 43 live voice conversations, the gate held 30/30 attempts to skip it, opened 10/10 when confirmed, and every repair line was spoken word-for-word in 40/40 completed runs.
• Built the real-time lane on Amazon Nova 2 Sonic (speech-to-speech) behind a custom relay that constrains tool calls, with 8 MCP tools (~2 ms each) secured by Cognito JWT auth on AgentCore — plus live recall checks against the US CPSC government API and DynamoDB-persisted household state.
• Designed an offline pipeline turning manual PDFs into structured, human-reviewed repair data (46 page-cited corrections across 5 appliances), and open-sourced cpsc-recall-check — an MIT-licensed Python library with 21 tests against real recall records.

Full-stack AI copilot for international students — paste a job link and upload a resume to get a one-page prep sheet: live company intel with cited sources, visa-sponsorship signals (H-1B/OPT/STEM OPT) with evidence and confidence, real interview questions, networking contacts, a fit score, an AI-tailored resume, and “Ray”, a chatbot grounded in all of that research.

• Cut LLM usage from ~10 to 2 calls per prep — running entirely within free tiers — via a semantic router that offloads factual queries to Tavily (zero LLM calls), a global semaphore with per-key rate limiting, 429/503-aware retries, and response caching.
• Architected a multi-provider AI pipeline: LangChain orchestration with structured Pydantic outputs from Gemini, plus Tavily running parallel, source-cited live web research so every fact in the app links to evidence.
• Shipped end-to-end to production with Google OAuth + Supabase Postgres for accounts and saved application history — React/Vite frontend on Vercel, FastAPI backend on Render, CI via GitHub auto-deploys.

Carrier-agnostic agentic loop that completes an 8-page workers' compensation insurance quote end-to-end with zero carrier-specific code — an OpenAI reasoning model drives a Playwright MCP server controlling a real Chromium browser via function calling.

• Engineered a self-healing core that auto-captures the live page snapshot on any tool failure and reinjects it, letting the model recover from complex Angular/PrimeNG widgets in real time without hard-coded selectors.
• Delivered a validated live submission ($539 instant quote, 0 errors, ~9-minute run); added recipe record-and-replay so routine runs cost cents — cutting projected cost at scale by ~10x.
• Designed a modular TypeScript architecture (agent loop, MCP client, per-carrier mapping, CLI) with token/cost tracking — adding a new carrier requires only a single new file.

Real-time LLM voice assistant for UNT's College of Merchandising, Hospitality and Tourism — helps students and faculty locate rooms and provides front-desk help.

• Built with Gemini + LangChain, modular tool-calling, and session memory for context-aware multi-turn workflows.
• Optimized API latency and token usage via prompt caching, batched requests, and efficient context window management.

Hands-free AI fitness companion using Whisper STT, React, n8n, and OpenAI. Recognized as Top 14 of 32 teams at Dallas AI 2025.

• Built a voice-enabled AI assistant with Whisper STT and a responsive React front-end.
• Designed automated task workflows in n8n reducing manual interaction steps by 60%+.

End-to-end GCP data pipeline processing 1,048,575 CDC BRFSS healthcare survey records (2020–2024) into real-time informatics insights.

• Engineered scalable ETL, batch, and streaming workflows with Cloud Storage, BigQuery, and Spark + Hive on Dataproc.
• Reduced health-alert delivery latency from minutes to under 5s via Pub/Sub streaming into BigQuery.
• Delivered Power BI dashboards across 27 BRFSS variables for near-real-time public health monitoring.

End-to-end batch pipeline ingesting 10M+ NYC Yellow Taxi trips into BigQuery with bronze/silver/gold medallion layering and dbt-modeled analytics.

• Orchestrated ingestion, transforms, and Great Expectations data quality checks as Apache Airflow DAGs with retries.
• Published metrics in a Looker Studio dashboard for at-a-glance KPI monitoring.

Browser-based multiplayer brawler for up to 6 players with three Brawl Stars-style modes — friends join from any phone or laptop with a 4-letter room code, no login, no install. Runs with zero game servers, entirely on free tiers.

• Architected serverless netcode: the host's browser runs the authoritative simulation at 30 ticks/s in a Web Worker while peers connect over WebRTC data channels (an unreliable lane for snapshots and inputs, a reliable lane for events), with Supabase Realtime used only for the handshake and a 15 Hz relay fallback.
• Implemented competitive-grade lag compensation — local prediction with reconciliation, remote players interpolated 80 ms in the past, shots fast-forwarded by measured latency, and host-side visibility culling so players hidden in bushes are never even sent to opponents.
• Built a code-only art pipeline: Blender Python scripts generate and export every model as glTF, rendered live in Three.js with soft shadows, animated water caustics and poison gas — with a 2D sprite fallback for devices without WebGL.