Pravina Bariya

GenAI/RAG assistant for Alfabet enterprise documentation (thesis project). Chose RAG over fine-tuning (documentation updated every few months); FAISS + BM25 with rank fusion and re-ranking. Nova Lite about 50× cheaper per token than Claude at comparable in-scope faithfulness (9 models, 453 records; thesis). 37 of 41 refusals were for answerable questions (single annotator). Tested by customers and integrated into Alfabet by end of 2025.

Documentation team (05/2022 – 01/2023); thesis research and prototype chatbot in the product management team (from 02/2023). Researched AWS Bedrock models, RAG versus fine-tuning, chunking and vector databases; built a standalone prototype chatbot with source tracing, tested by the internal team by April 2024. The prototype became the starting point for the Bizzdesign assistant, after Bizzdesign took over the Alfabet product line (new contract from 01/2025).

Final grade 2.5 (gut); thesis grade 1.7. Master's thesis: Developing a Domain-Specific AI Assistant for Technical Documentation Using AWS Bedrock Models (03/2026); part 2 tests the same architecture on a second corpus, Berufearchiv (github.com/Pbariya/berufearchiv-rag). Completed part time alongside working-student roles.