AI-Driven Conversational Rental Marketplace with Fuzzy Hybrid Search and Trust-Based Moderation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Finding rental accommodation in the Kathmandu Valley isfragmented, broker-dependent, and plagued by fraudulent listings.Conventional property portals reduce rich, hyper-local requirements—water reliability, sunlight, power backup, neighborhood character—to a handful of rigid filters, and offer little defense against fakeadvertisements. This paper presents RoomieKtm, a domain-specificrental-marketplace system that integrates conversational search,hybrid retrieval, fuzzy re-ranking, and automated trust scoring—eachbuilt on established techniques—into a single, purpose-built pipelinefor this market. We describe a hybrid retrieval path that fuses lexical(BM25) and dense semantic (vector k-NN) candidates via ReciprocalRank Fusion, then re-ranks them with a weighted, zero-order Sugenotypefuzzy aggregation that models graded, tolerant preferences. Tocombat fraud, we introduce a five-factor trust-score model that drivesa three-way automated moderation router, complemented by a humanadmin review queue, tenant reporting, and an append-only audit log.We further describe an administrator-reviewed landlord identityverification(KYC) workflow with explicit data-protectionsafeguards. We evaluate the pipeline on a controlled pilot collection(93 listings, 42 labeled queries) with a four-arm ablation—BM25,semantic-only, hybrid without fuzzy re-ranking, and the full hybridplus-fuzzy pipeline—and report paired significance tests alongsidepoint estimates. The fuzzy layer recovers ranking quality specificallyon paraphrased, vocabulary-mismatched queries (nDCG@10 0.662vs. 0.622 for RRF fusion alone), and the trust router correctly flags 13of 15 constructed fraud patterns (recall 0.867, accuracy 0.914);however, none of the observed differences reach statisticalsignificance at this sample size (Wilcoxon signed-rank, all p>0.05), alimitation we report plainly rather than obscure. We treat this as anhonest, small-scale validation of correct implementation and designdirection, not a proof of production-scale performance, and we detailexactly what a larger, real-world evaluation would need to add.