Mounika Bharatha

A RAG search engine over 200 federal grant PDFs (15K chunks), with section-aware chunking, metadata filters, and hybrid pgvector + full-text search merged by reciprocal rank fusion.

• 82% → 90% recall@5 after adding a cross-encoder reranker; section-aware chunks beat fixed-size (82% vs 63%)
• 97% faithful answers built only from word-for-word verified quotes, with PDF page citations
• 13 of 14 unanswerable questions correctly refused on a 93-question eval set
• $0/month as a TypeScript Cloudflare Worker on Neon Postgres; CI fails the build if recall drops

A full-duplex voice agent for behavioral interview practice. It asks 3 questions from a bank of 14, follows up when an answer is thin, and scores feedback on 5 competencies. Browser audio streams into NVIDIA Nemotron ASR, Qwen3 on Groq and NVIDIA Magpie TTS.

• ~3s → ~1.5s voice-to-voice, via sentence-by-sentence synthesis and prewarming
• 1.3–2.6s adaptive end-of-turn detection that tolerates thinking pauses
• ~300ms barge-in that cancels the reply when you interrupt
• 42 of 70 ungrounded quotes rejected by a transcript check · 66 tests

A multi-step agent that splits a question into sub-queries, retrieves live web results with Tavily and writes a cited answer, with a bounded self-critique loop that searches again for gaps.

• RAG memory past reports embedded with gemini-embedding-001 (3072-dim) in a SQLite vector store; top 3 matches feed the planner
• 0.80 cutoff set after measuring that unrelated questions still scored ~0.70
• Pause & resume with checkpointing, plus human approval before anything is finalized

A multi-page website built and deployed for a real bakery and boba shop. Customers can place pickup orders through a serverless ordering flow with Resend email confirmations, and the site ships with JSON-LD structured data for search.