Sri Harshitha Yannam

About

LLM/agent engineer who shipped a production "Memory Assistant" at HydroX AI, building a LangChain/LlamaIndex RAG memory pipeline on ChromaDB/FAISS with robust fallbacks (BERT/BART), prompt-injection mitigation, and 99.9% uptime monitoring. Also built a multi-step customer support agent using Rasa + OpenAI Assistants API with structured tool calling, guardrails, and human-in-the-loop escalation, and has experience hardening agents against messy ERP data via Pydantic validation, idempotency, and transactional outbox patterns.

Key Strengths

• Shipped production LLM memory assistant with long-term context retention and multi-turn session management
• Built RAG pipeline with summarization + metadata capture into ChromaDB/FAISS for real-time historical context retrieval
• Designed reliability framework: 4xx/5xx classification, automated fallbacks, prompt-injection risk mitigation, SLA monitoring to 99.9% uptime
• Implemented BERT/BART fallback path to reduce hallucinations when retrieval relevance is low
• Solved 'lost in the middle' via cross-encoder re-ranking; improved retrieval precision/accuracy by 35%
• Optimized latency by parallelizing vector search and intent classification with Python asyncio
• Architected multi-step customer support agent with Rasa state-machine orchestration and OpenAI Assistants API tool calling with structured JSON
• Implemented guardrails (intent confidence thresholds, business validators) and human-in-the-loop escalation for high-risk/looping cases
• Handled messy ERP-like data with Pydantic schema enforcement, idempotent UPSERTs, and transactional outbox pattern for partial failures
• Reported business impact: 68% increase in session management efficiency, 25% reduction in query resolution time, 45% boost in user satisfaction

Experience

Education

Skills

Languages