Ojasmitha Pedirappagari

About

Built and shipped a production RAG-based assistant that lets parents ask natural-language questions about their child’s learning progress, using pgvector retrieval (child-id filtered) and Redis caching to hit ~180ms latency. Implemented real-world guardrails and compliance (Llama Guard, COPPA, retrieval thresholds, fallbacks) with 99.5% uptime, and ran human-in-the-loop eval loops that improved satisfaction from 3.8 to 4.2 while serving 60k+ monthly users and reducing costs significantly.

Experience

Education

Key Strengths

• Shipped production LLM/RAG assistant end-to-end for parent Q&A on child learning progress
• Strong performance tuning: reduced response latency to ~180ms by scoping retrieval (retrieval scope vs performance tradeoff)
• Built production guardrails and compliance controls (Llama Guard, COPPA, retrieval thresholds, fallbacks) achieving 99.5% uptime
• Designed and ran LLM evaluation loops using real user queries and multi-metric scoring; improved satisfaction average from 3.8 to 4.2
• Scaled LLM/data pipeline to 60k+ monthly users with observability (GCP logs/dashboards) and reliability patterns (retries, exponential backoff)
• Cost/latency tradeoff execution: increased latency to ~45s to cut monthly cost from ~$500 to ~$180 while maintaining accuracy