At Grant Thornton AG, I designed and benchmarked eight LLM orchestration architectures for financial-statement mapping, comparing them against labeled data. I developed a reusable multi-agent framework with typed contracts, model routing, and feedback-driven evaluation.
I built a document-grounded RAG assistant with PDF and DOCX ingestion, vector retrieval, source citations, and an agent tool. I also exposed its asynchronous core through web, FastAPI/SSE, and Microsoft Teams.
I engineered a production FastAPI and Azure Queue/Blob processing platform with KEDA-scaled workers, job isolation, and stalled-job detection. Direct-to-Blob transfer and chunked processing raised supported uploads from 180 MB to 5 GB.
