The traditional AHP (Analytic Hierarchy Process) algorithm uses the judgment matrix generated by the subjective judgment of experts, which makes it difficult to ensure the consistency of the judgment matrix, and the weight distribution is highly subjective, so the assessment results are not objective enough. This paper uses the RF (Random Forest) algorithm and triangular fuzzy numbers to improve the AHP algorithm and optimize its application in the fire safety risk assessment of construction engineering. The study considers the characteristics of construction engineering and fire safety characteristics to establish an accurate indicator assessment system, and assesses the safety risk situation from the aspects of building fire protection design, fire protection facility settings, fire safety management and control, hazard sources, and rescue conditions. Then, triangular fuzzy numbers are used to replace traditional exact values to process uncertain data, and the data-driven RF algorithm is used to identify the importance of indicators and construct the fuzzy judgment matrix of AHP. Finally, the improved AHP algorithm is applied to the fire safety risk assessment of a construction engineering in Beijing, and the risk assessment results are output. The experiment takes office buildings in construction companies in Beijing as the main research object. The results show that the number of instances of risk level assessed by the improved AHP algorithm (RF + triangular fuzzy number) is consistent with the actual one; the instance data assessed by the traditional AHP algorithm has the largest difference with the actual one, reaching a maximum of 5, and the CC index reaches 0.975. The experimental results show that this paper applies RF and triangular fuzzy numbers to improve the AHP algorithm, which can greatly ensure the objectivity of weight distribution and the consistency of judgment matrix, thereby making the fire safety risk assessment of construction engineering more in line with the actual situation.