Problem
Institutional research teams need dependable access to large, heterogeneous knowledge collections without relying on ungrounded model responses.
Context
The current site describes this work as a full-stack LLM assistant for retrieval, question answering, and knowledge automation.
My role
Led development of retrieval, reranking, agentic workflow, and evaluation components.
Constraints
- Retrieval quality must remain inspectable.
- Responses need grounding and safety checks.
- The system must support maintainable, production-oriented workflows.
Architecture
The source material identifies hybrid retrieval, metadata indexing, cross-encoder reranking, task decomposition, memory modules, orchestration, and automated evaluation.
Technical decisions
- Combined retrieval methods before reranking.
- Added explicit workflow orchestration for multi-step tasks.
- Evaluated hallucination, relevance, and safety behavior.
Trade-offs
Document latency, retrieval depth, model selection, hosting, and cost trade-offs.
Results
The project demonstrates an applied architecture for dependable research knowledge access, emphasizing grounded answers, retrieval inspection, and evaluation loops.
Public note
Public materials are limited to a high-level architecture summary because the system is tied to institutional knowledge workflows.
Related links
Related public references can be listed as institutional sharing rules allow.