oopscompiled

Architected and deployed an enterprise-scale RAG (Retrieval-Augmented Generation) system for internal knowledge management using Llama 3.2 and Qdrant, reducing average employee information retrieval time by 65%. Fine-tuned the LLM using LoRA (Low-Rank Adaptation) on proprietary internal datasets, significantly improving domain-specific reasoning and aligning model output with strict corporate compliance guidelines. Engineered a high-performance polyglot architecture, utilizing Python strictly for ETL/Encoders and migrating the core inference engine to Rust (Candle). Achieved a 3x reduction in latency (220ms → 65ms) and reduced peak memory usage by 40% through consistent streaming without Garbage Collection lags. Built a hybrid search pipeline (dense + sparse embeddings) using LangChain and Qdrant, increasing the relevance accuracy of generated answers to 94%. Integrated the solution across HR and Tech departments, resulting in a 30% reduction in onboarding time for new hires.

Engineered a predictive model to forecast material needs, improving forecast accuracy from 78% to 85%. Tackled overstocking and shortage issues, reducing operational costs and avoiding project delays. Developed an integrated vendor-sourcing module to identify and recommend optimal local suppliers based on predicted demand. Proactively analyzed and resolved a critical vulnerability in backup system: discovered that models based on pandemic-era data frequently and significantly underestimated demand during holiday periods, thus avoiding the risk of large-scale shortages.