Prakhar Kumar Singh

Aurigo Software Technologies

• Engineered production multi-tenant AI chatbot on AWS Bedrock (Claude Sonnet 4.5) using Python FastAPI and WebSocket streaming - 250+ concurrent connections across 50+ enterprise tenants with 2-layer caching (Valkey + semantic) cutting latency by 60% and API costs by 40%
• Built hybrid RAG combining pgvector (1024-dim Titan embeddings) + BM25 via Reciprocal Rank Fusion (RRF), indexing 100K+ chunks in Aurora PostgreSQL - 85%+ retrieval relevance with LLM-based reranking
• Architected SQS-based async ETL ingesting 1000+ docs/day via PyMuPDF and Docling OCR with ThreadPoolExecutor (12 parallel workers) and adaptive memory scaling - 99.5% ingestion success rate
• Built NL-to-SQL engine with circuit breaker, Sonnet-to-Haiku failover, sqlglot AST validation for SQL injection prevention, and PII filtering - 500+ queries/day at 99.9% uptime
• Established RAGAS-aligned LLM evaluation framework (MRR, correctness, groundedness, completeness) with golden test datasets from 100K+ chunks and SQS-based regression testing for continuous drift detection
• Built TensorFlow/Keras DNN for infrastructure budget forecasting across 5000+ projects - 94% R², 40% error reduction, deployed on Lambda at sub-500ms latency with S3 model storage with SSM Parameter Store versioning
• Orchestrated 40+ table MWAA Airflow ETL from S3 CSVs and Pendo API into Aurora PostgreSQL across 10+ tenants - 85% failure reduction via intelligent CSV chunking and 30 concurrent parallel batch operations
• Established S3-based data quality monitoring with config-driven transformations - 99% data accuracy across 10+ customer instances in production

Aurigo Software Technologies

• Built PoC multi-document search chatbot using AWS Kendra and Anthropic Claude - validated retrieval quality on construction project documents
• Prototyped early RAG-based chatbot for construction document and database search, laying the foundation for the production multi-tenant system
• Developed sentiment analysis pipeline for construction bidding documents using transformer-based models
• Built early budget prediction prototype in TensorFlow/Keras - later evolved into the production forecasting system (94% R²)
• Contributed to multi-tenant S3/Pendo → Aurora PostgreSQL data warehouse ETL on AWS MWAA
• Converted to full-time Software Engineer in June 2024 based on performance