Vicky Jaiswal

AI Model Router

Classifies any query into 5 task types using Gemini Flash and routes it to the best free LLM via OpenRouter. Achieved 100% routing accuracy on a 27-query benchmark with confidence scores tracked per category. Includes LLM-as-Judge quality scoring and a Streamlit dashboard with Compare Mode, analytics, and full SQLite observability.

Semantic Search & RAG System

Production-grade 7-layer RAG pipeline for e-commerce semantic search using FAISS and SentenceTransformers. Implements Out-of-Domain detection, confidence-based rejection gates, and zero-hallucination guarantees — LLM generation is strictly grounded to retrieved products only. Includes Recall@K and MRR retrieval evaluation with full explainability logging.

AI Data Quality & Drift Monitor

ML monitoring pipeline that detects statistical drift using KS-tests and IQR-based anomaly detection with 96% accuracy. Extended to track LLM output quality via Jensen-Shannon divergence on sentence embeddings and ROUGE-L scoring. Automated retraining recommendations reduced undetected quality issues by 89%.