Ramya Sree Kanijam

About

Built and shipped a production LLM-powered RAG system at Upstart enabling natural-language search across 50k+ scattered internal technical docs. Delivered sub-300ms p95 latency for ~50 active users with strong hallucination safeguards (retrieval-first, thresholds, citations) plus robust testing/monitoring and cost controls (prompt caching cutting API spend ~20%).

Experience

Education

Key Strengths

• Built and deployed production LLM RAG search over 50k+ internal technical documents
• Achieved ~300ms p95 end-to-end latency at scale (~50 active users) via profiling, batching, parallelization, and ECS warm starts
• Hallucination reduction using retrieval-first design, similarity thresholds, and citation/"I don't know" prompting
• Production-grade testing approach (unit/integration/E2E) with ~75% coverage
• Strong observability and reliability practices (Grafana/Prometheus monitoring, regression detection)
• Cost optimization (prompt caching) reducing LLM API costs by ~20% while maintaining quality
• Data-driven iteration using offline evals and production metrics (latency, confidence, error rates, user feedback) including A/B tests
• Effective collaboration with non-technical stakeholders using demos and metrics to translate needs into requirements
• Designed and owned end-to-end perception-to-decision backend pipeline for robotics
• Reliability engineering for noisy/real-world sensor edge cases (validation, buffering, timestamp alignment)
• Built deterministic debugging workflows using structured logging and replayable tests (e.g., rosbag-style)
• Systematic real-time debugging: reproduced issues from logs, validated TF timing, profiled latency, tuned loop rates and dropped stale frames
• Strong integration focus across distributed/heterogeneous robotic components with strict interfaces and failure-safe behavior