Anshikabajpai

Led design and deployment of a production hybrid RAG + rules-based AI system across multiple enterprise data sources and live clients. Resolved cross-domain query ambiguity using confidence thresholding and rule-based fallback routing. Built LLM extraction pipelines converting unstructured documents into structured data with ~98% field-level accuracy. Designed LLM evaluation and risk-gating layers (LangSmith/OpenEvals) to catch low-confidence outputs pre-production (95%+ coverage, ~30% reduction in unreliable outputs), achieving 84% factual accuracy on the production evaluation set.

Built a GAN-based unpaired image-to-image translation pipeline on NVIDIA A100 GPUs, reducing cross-domain distribution mismatch by 37% while preserving structural detail and enabling training on unlabeled multi-modal imaging data. Trained a CNN-based segmentation model on 11,000+ unlabeled samples using translated data, improving Dice score from 43% to 78% via targeted loss weighting, data augmentation, and learning-rate tuning—reducing the need for manual labeling.

Built an end-to-end RAG pipeline on GCP using Vertex AI embeddings and Enterprise Search with semantic chunking and Redis caching. Reduced inference latency by 75% (4.8s to 1.2s) and redundant API calls by 41%, improving Precision@3 by 18% on cybersecurity queries via A/B testing. Quantized a model to TensorRT FP16 using ONNX to optimize low-latency inference. Architected an asynchronous ingestion pipeline (GCP Pub/Sub) for real-time PDF/DOCX/HTML processing with BigQuery content-hash based deduplication, reducing duplicate ingestion by ~35% across 500+ sources. Engineered a multi-turn conversational agent using context compression and sliding-window memory to support 20+ turns, improving answer relevance by 31% using chain-of-thought and few-shot prompting.