SQL Auto-Generation Large Model PlatformResponsible for training the SQL large model and prompt engineering based on the DIN-SQL framework. 1. Data preparation includes self-structured and manually annotated data, as well as open-source data (Spider). The data format consists of (Question, Schema_links, Step_by_Step, final SQL). Route prompts include general questions, easy prompts (simple SQL without join), mid prompts (two-table join), and hard prompts (multiple table join).2.Model training involves using CodeLLama34b for LoRA training, covering tasks such as route classification, SQL generation, and general text generation.3. Model deployment utilizes TGI and VLLM frameworks for routing user questions (non-SQL/easy/mid/hard), generating schema_links (containing fields and join relationships used in SQL), and finally incorporating the generated results into prompts to produce the final SQL

HuiXing cross-recommendation system: Optimize the revenue and user conversion rate in the transit route, use the combination of XGB and deep learning to solve the pain points of users' travel difficulties during peak periods and low success rate of ticket grabbing, and recommend various types train plus train etc. for users under different search categories. Combination of multiple routes such as train + plane, increase user click rate and order rate;Artificial Intelligence Customer Service: Through user intention identification and understanding, word segmentation and sentiment analysis are performed based on basic features in historical data. For single-round QA and multiple tasks, build an OTA industry knowledge and label system, use FastText and Bert offline training to achieve an accuracy rate of more than 90%, perform Query retrieval and recall for different user intentions, and pass the possibility of candidate intentions. Higher reordering fusion rules increase the matching of cold starts and high frequency problems;UGC Community building: Build the UGC content platform of Tongcheng Mini Program from scratch, start from tag annotation and user travel records, identify travel boutique stickers in the early stage and distribute them to the corresponding city and age group user groups, dynamically calculate user interests through MAB and recall the most suitable content, use the generalization index of each stage to train the DNN model under multi-objective to solve the content distribution and sorting;

Add user imei identification, verification code page operation and click behavior, model matching, IP fluctuation positioning, GPS black product aggregation, startup and running time, add unsupervised model on the basis of lightgbm, filter features based on varclustering, and generate iforrest tree model Identify the results, intercept 20-40% of registered black products every day;Create a fake coupon transaction model and a dynamic coupon issuance model. There are two types of models: new customers and regular customers. The regular customer model is mainly used to predict whether to place an order again. Confirm the amount of coupons issued for thenext event and whether there is a threshold. The bottom layer of the two types of models uses LSTM and XGB to output normalized scores, and the upper layer uses LR to improve interpretability. After going online, increase the conversion rate of orders by 10%, and the order rate of customers who issue coupons is high;According to the ResNet model, it only costs one camera in each hotel to monitor and identify the number of people who deliver ID cards backand forth at the hotel front desk every day in real time, to compare the number of orders entered by hotel owners on the PMS platform, and to further accurately identify false hotel transactions level, and make reports to visualize and analyze abnormal fluctuations of hotel OCC indicators;

Recommendation ranking: Using the user's app installation list and the demand preferences of different loan products, according to the historical user's click habits, using clustering and classification algorithms, depicting different portraits of users to improve the user's product purchase conversion rate by 20%;Social network: For the first time in our company, we use the app tracking point and user operator data to code 0 and 1 for users through the data funnel method, and then add the A-card score. Each time, the user's social score is updated according to the weight of the social network. User social risk system, and add B card and C card as features, about iv0.1;Warehouse Building: Responsible for the design, construction, optimization and integration of risk control data warehouse, including demographic characteristics, behavior data and loan data, among 10 tables with over 2000+ fields. Greatly improve the modeling efficiency;Anti-fraud model: Responsible for the anti-fraud project of China Everbright Bank, building process with createing a total of 15 wide hive views. Through preliminary analysis and cross-validation, 10+ feature multi-dimensional subjects are designed, 5000+ derived variables are created through FRM clustering, ylabel is established semi-supervised, and text-type features are classified;

SQL fetching, variable ETLization, variable screening, Python development of logistic regression model, a total of 5 versions of the model were established, including pre-loan and post-loan scorecards. The repayment rate is above 91%;Mining nearly 20 valid strong tags from third-party data sources, user text messages, and bank flow, and putting them into the pre-loan scorecard model, the overall AUC of the pre-loan scorecard is increased by 1%, and the ks value is increased by 2%;

Trade on the second market for stocks

Foucs on the basic bank business, customer guidance with simple management and sales promotion of finanicial products.