2.5 million light years

Smarter Supplier Recommendations with LLMs? Sure thing, but wait.. - Can existing recommendation systems fully understand supplier fit for complex service sourcing events just because it's armed with LLM? - Can web-scale supplier intelligence improve candidate discovery? - Can LLMs reason better about supplier relevance than purely ML-based ranking models? - And what happens if we combine all of the above and apply it to legacy clients? During the hackathon, we explored these questions by building a hybrid AI recommendation system combining: - LLM-powered supplier web search - Existing recommendation candidates from our ML stack - LLM-based ranking and reasoning on top of supplier candidates The goal: improve supplier relevance and increase buyer confidence in service sourcing events. Early Results are bitter-sweet but promising! Our experiments showed encouraging signals across all three hypotheses: A. LLM-powered supplier web search surfaced relevant and previously unseen suppliers B. LLM-based ranking often improved recommendation accuracy over the baseline C. The combined setup demonstrated noticeable quality improvements for service sourcing use cases compared to Andromeda 1.0 At the same time, the experiments exposed several important limitations: - LLM ranking quality is still highly unstable - LLM demonstrated strong execution bias, attempting to “solve” tasks by any means, often over-processing requests - LLM ranker struggled to autonomously identify and prioritize the most meaningful evaluation signals The hackathon confirmed something important for us: LLMs are not yet a replacement for recommendation systems, but they are becoming a very powerful layer on top of existing ML infrastructure. Git: https://gitlab.fmdev.io/ds-team/hackf27-2-5-million-light-years