Built a video-call app where each person speaks Hindi, Telugu or English and hears the others translated into their own language, using real-time speech-to-speech models (Google Gemini Live Translate, OpenAI GPT-Realtime-2 and GPT-Realtime-Translate, chosen per meeting). Translation runs in the speaker's browser and the translated voice is swapped into the WebRTC call, with live captions relayed over Socket.IO. Compared all three models on real two-person calls and wrote an interpreter prompt that stopped GPT-Realtime-2 from answering questions instead of translating them. Deployed on Vercel and Render with a TURN relay for calls across different networks.

Built a RAG app that indexes a user's top 15 GitHub READMEs into Redis vector search, so questions about any repository are answered from that repository's own text. Cut embedding cost with 1,500-character overlapping chunks, batched Gemini embedding calls (32 texts across 5 workers) and a 24-hour SHA cache that skips unchanged READMEs. FastAPI backend with PostgreSQL for accounts, OTP email sign-up and background indexing jobs.

Designed a conditional Retrieval-Augmented Generation (RAG) workflow using LangGraph that intelligently routes user queries to either direct LLM inference or document retrieval based on query intent. Built a semantic retrieval pipeline by processing PDF documents into vector embeddings, indexing them with FAISS, and retrieving relevant context using LangChain for grounded response generation. Developed an interactive Streamlit application integrated with Groq LLM to deliver low-latency, context-aware responses with modular workflow orchestration for extensibility.