At Citimedia Global Solutions Private Limited, I moved a product from a paid LLM vendor to self-hosted Gemma 4 12B on NVIDIA L4 with vLLM, eliminating per-token API spend while maintaining 258ms p95 time-to-first-token at 8-way concurrency. I also shipped a real-time voice agent with custom turn-taking and barge-in, opening 11 languages and cutting time-to-first-audio by 21%.
I designed a multi-tenant RAG pipeline for a Marathi/English government assistant, raising recall@5 from 69.2% to 92.3% with hybrid retrieval on pgvector. At Aarambh AI, I launched its first real-time clinic voice assistant in English and Telugu and built a document-to-JSON pipeline across 10+ specifications.
