I designed the platform’s knowledge system as infrastructure, not a prompt: its own retrieval logic, its own ingestion pipeline, and its own way for non-engineers to extend it.

• Designed a retrieval strategy with explicit selection logic: score-gap selection, an LLM tie-breaker, and human handoff when confidence is low. Ran comparative experiments between LLM-based and bi-encoder rerankers (e.g. Cohere) across chunking strategies, and chose the LLM-based reranker — it cost more on latency and compute than the bi-encoder option, but correctness mattered more for this use case.
• Built a policy-document ingestion pipeline on multiple PDF parsers with a human-review step before documents enter the tenant-aware embedding store, running chunking-strategy experiments alongside the reranker work to tune retrieval quality end-to-end.
• Built a connector framework letting community managers configure their own knowledge sources — policy documents, web search, third-party APIs, property websites — shipped and in production use.
• Backed the flow with a ~200-question evaluation dataset and an LLM judge, built alongside the feature rather than bolted on after.
• Deliberately deferred AI-generated knowledge-base answers from the MVP until validated.