dukeyuan

1. Participate in the development of agents to implement data retrieval based on user queries, design tools, select templates, and display data. 2. Design a technical solution for UI generation using large models; fine-tune and test the UI interface generation effects using qwen2-7B and InternVL-8B. 3. Train a CLIP model to calculate the similarity between text and images; used to evaluate the similarity of images recalled based on RAG and the query. 4. Train a text-to-text 1B model and test its deployment capability on mobile devices. 5. Participate in the development of multimodal large models for in-car application scenarios; develop OpenAI interfaces based on the InternVL-8B quantized model compiled with trt_llm, aligning with OpenAI services.

1. Based on the GPT 2 model, ChatGLM and LLAMA models use the company's popular science articles, medical data, and train GPT to generate models for medical GPT Q&A. 2. Complete the Reward_ Train the model and then complete RLHF training. 3. Train multiple models to determine the GPT generation results and reduce the generated hallucinations. 4. Based on the GPT 2 model, use the company's medical chat data to train the consultation GPT model for multiple rounds of medical consultation dialogue training. 5. Conduct research on major model solutions such as ChatGLM and Langchain, and test the effectiveness of medical data. 6. Establish a user profile system; This includes static tags such as the user's age, gender, location, family members, and my doctor, as well as dynamic behavior of the user, including related information such as search, browsing, consultation, purchase, and shopping cart. 7. Utilize OCR to identify patient prescription information, and greatly improve the entry efficiency of the store by extracting and structuring case information. 8. Generate a vector of drugs based on the doctor's prescription, and obtain a vector of diseases based on the relationship between diseases and drugs; Based on the relationship between doctors and diseases, obtain doctors' vectors, calculate the correlation between drugs, diseases, and doctors based on drugs, diseases, and doctors' vectors, and store them in neo4j to form a knowledge graph.

October 2018 to January 2022Lead a team of 10 people to implement recommendation and search system from 0 to 1 Programming accounts for 75%, and team management accounts for 25% • Homepage course recommendation: 1. Based on the business scenario of Chaos University, analyze and transform the actual scenario of education intelligence; 2. Use user behavior data to analyze and model, make personalized course recommendation, and continuously evaluate and optimize the model; 3. Provide decision-making for product optimization based on data analysis; 4. Lead the team to build the course recommendation system from 0 to 1, build the course portrait, user portrait, offline training samples, offline model training and other modules; 5. After the recommendation system goes online, the pv and uv will increase by more than 20%, the average number of clicks will increase by more than 10%, and the viewing time of users will increase by more than 30%; 6. Update the real-time user profile, build the real-time recommendation system, build the user diversion system, and track the effect of the recommendation system; 7. Build a recommendation system evaluation system; 8. The recommended model is tested by GBDT, GBDT+LR, FM, DNN, wide&deep, DCN and other models; • Home app search: 1. Responsible for designing the search, sorting and implementation of app homepage; 2. Personalization is realized inside the sorting, which is more than 15% higher than the original search effect; 3. The memory-based data storage method greatly improves the search efficiency, and the average request time is 0.1s; 4. Search includes course search, note search, excellent course search, topic search, short video search and other modules; 5. Word segmentation, intention recognition, intelligent matching and other methods are used in search to reduce the no-result rate by more than 20%; • Recommended courses: 1. Responsible for the design, sorting and implementation of relevant courses; 2. Use the title, sku, teacher's name, article keywords and other dimensions to evaluate; 3. More than 50% higher than the original effect; • Classroom practice recommendation: 1. Be responsible for the design, sorting and implementation of the recommended practice in the classmate circle; 2. Use the title, author's name, topic keywords, pictures in practice, user comments and other dimensions to evaluate; 3. More than 20% higher than the original effect;

Mainly responsible for the personalized news recommendation project of Tencent WeChat plug-in, Tencent News, Daily Express and Tencent Smart Push• Natural language processing module: The maintenance of thesaurus, through the discovery of new words, continues to maintain the company's corpus, million level; Word segmentation, based on jieba , calculates the idf of words, marks the part of speech, based on babbie word segmentation, uses word vector to do synonym mapping, reconstructs word segmentation, and keyword extraction algorithm; Classification, based on LDA algorithm (reconstruction), carries out unsupervised clustering to provide corpus basis for unlabeled classification; Using textcnn, naive Bayes (reconstruction), compatible with sparse data, and the accuracy of primary and secondary classification is more than 95%; Place name extraction: based on the place name corpus, extract the article place name; Using word2vec to generate word embedding; • User portrait module: Based on the user's click behavior, build the user tag system through the news tag system; Based on the time series, the user interest in the long and short term is constructed and attenuated; Based on GPS acquisition, obtain user's activity track and judge user's static attribute; Based on the embedding results of words, build the user's embedding; • Recall sorting module: Based on the first-level, second-level and third-level label system, recall is carried out according to the user portrait label, news popularity and operation strategy; Recall based on news embedding similarity; Based on DMLC, build a distributed training system and training model to support the training of 100 million parameters; Based on the embedding results, carry out depth models such as GBDT, GBDT+LR, DNN, wide&deep, and tree models for training prediction; Mining long and short term features, matching features, similarity features and other training models; • Engineering technology: According to the size of data increment, build a news pool to realize the online and offline of data; Channel app recommendation, providing recommendation services for all sub-channels; Based on mongodb, build a real-time updated user portrait service to serve the entire recommendation system; Build nlp service to process various label extraction in real time; Based on the parameter server, the distributed learning system DMLC is developed and used to support the research and development and optimization of a high-performance distributed training platform with hundreds of billions of features and trillions of parameters in high-dimensional sparse scenes, greatly improving the efficiency of offline training evaluation, and improving the training speed by 3-5 times;

Mainly responsible for the research of pattern recognition and license plate location algorithm: 1. Analyze the paper and prepare the feasibility study report;2. Write a program (C++) according to the actual problem to verify and process; 3. Write a program and use opencv to locate the license plate position in the image according to the video image data; 4. Be responsible for the transplantation of software algorithms and software as a whole, and the realization of functions; 5. Be responsible for the combination of algorithm and software, making dll files, and realizing dynamic links; • Voice: 1. Be responsible for the performance tuning of speaker segmentation; 2. Understand the relevant knowledge of vad, be responsible for the optimization of algorithms in vad and the selection of speaker segmentation; 3. Rewrite the algorithm for calculating pitch; 4. Understand the relevant knowledge of speaker segmentation and clustering, and optimize and improve the algorithm for specific data; 5. Test the improved algorithm and write a test report; 6. Write programs to realize other functions of the system. 7. Complete other tasks arranged by the leader. • Recommendation: 1. Build the Hadoop operating environment and build the Hadoop platform; 2. Research the mahout recommendation system, be responsible for the establishment of the recommendation system, and carry out the recommendation test; 3. Statistics of user logs and extraction of user log information; 4. Build a database to store user log information; 5. Research the word segmentation service, and modify and optimize the extraction of place names; 6. Provide web services, support the work of other departments, and provide segmentation and data query services for other departments; 7. Extract user browsing records and analyze them to extract user