SINCE 2025
Perplexity
AI/ML Engineer
- Production RAG Grounding
- Hybrid Retrieval Pipelines
- LLM Inference at Scale
- Document AI to LLM Systems
AI/ML Engineer · Perplexity · Ex-NVIDIA
Yamini Ravuri
United States
The engineer who makes research-grade models hold up under production-scale traffic.
Yamini Ravuri is an AI/ML Engineer at Perplexity, where she builds the hybrid Vespa–FAISS–BM25 retrieval and citation-grounding stack that lifted answer relevance 32% for 3,000+ daily active users. She came up through document AI at Accenture in India, then LLM inference at NVIDIA, where TensorRT-LLM pipelines cut serving latency 38%. What is distinctly hers: she engineers factual grounding as a measured subsystem, building hallucination-detection frameworks on DeepEval and RAG evaluation pipelines.
- 32%Answer relevance gain at Perplexity
- 41%Token throughput gain at NVIDIA
- 38%Response latency cut at NVIDIA
- 29%Inference cost cut at NVIDIA
HAPPENING NOW
STANDING TODAY
HOW THEY THINK4 BELIEFS
Four beliefs run under five years of production ML.
THE IMPACT
Five years of production systems — pull one out.
EVERYTHING ELSE · 5 WORKS
THE TIMELINE
The same problem — find the right information and serve it at scale — raised one level with every move.
The whole record, newest first – tap any moment and Aster tells the story.
2025-PRESENT
Grounding the answer, cited and fast
At Perplexity, Yamini builds the hybrid retrieval and grounding stack behind AI search — Vespa, FAISS, and BM25 fused into one pipeline that lifted answer relevance 32% for 3,000+ daily active users. Multi-stage reranking with BERT cross-encoders and XGBoost cut irrelevant results 41%, and a DeepEval-backed citation-grounding framework checks that every claim traces to its evidence.
“Always open to connecting with others building production AI systems. Let's talk retrieval, inference, or anything in between.”
2023-2025
Making large models serve fast and cheap
The technical hinge of Yamini's career. At NVIDIA she engineered distributed LLM serving on TensorRT-LLM, Triton, and Kubernetes, cutting response latency 38% and raising token throughput 41% with continuous batching and quantization. She completed her MS at Texas A&M–Kingsville in the same window, finishing the degree while already working full-time, and her FAISS and LangChain retrieval workflows here seeded the grounding work that followed.
“I specialize in LLM inference optimization, semantic search, and retrieval pipelines turning research-grade models into low-latency, reliable platforms that serve thousands of users.”
2020-2023
Learning production ML on enterprise documents
Yamini's longest role and the origin of her craft. At Accenture in India she built intelligent document processing for insurance and banking — BERT entity extraction and LayoutLM layout-aware pipelines that improved field-level accuracy 32%, plus OCR microservices ingesting millions of documents monthly. The production discipline she learned here — microservices, MLOps retraining, monitoring — carried straight into her LLM work.