Fraction concepts are challenging for many elementary school students because they require coordination among symbolic, visual, and contextual representations. This study developed and conducted a preliminary evaluation of a large-language-model-based mathematics chatbot that incorporates selected Dayak Paser cultural contexts for fraction learning in Balikpapan. The study used a Research and Development approach organized with the ADDIE framework and a one-group pretest-posttest evaluation involving 60 fourth-grade students. Evidence comprised expert product validation, student responses, pretest-posttest scores, the System Usability Scale (SUS), a descriptive review of 120 chatbot answers, and latency records from 100 interactions. The product-validation dimensions were rated 92%-94%, with an overall score of 93.4%. Positive student-response percentages ranged from 92% to 95%, and the SUS score was 87.6. Mean test scores were 55.3 before and 84.7 after use, a within-group difference of 29.4 points. The normalized gain calculated from the reported group means was 0.66 (moderate). The reported paired-samples result was t(59)=18.93, p<.001; the corresponding paired standardized change was dz=2.44 and the reconstructed 95% confidence interval for the mean difference was 26.29-32.51. These findings describe a substantial pre-post change but do not establish that the chatbot caused it because the design did not include a comparison group, randomization, a retention test, or a documented intervention dose. The study therefore supports product feasibility and warrants a controlled follow-up, while making no empirical claim regarding adaptive learning, increased motivation, cultural preservation, or retrieval-augmented generation