Jerry AI tutor
No reviews yet

How can machine learning help you at work and study :
1. Bank Risk Control application: Credit Score Card and anti-fraud Model
2. E-commerce scenarios: CTR(click-through rate estimate) in search ranking and recommendation
3. Computer Vision: OCR and face detection, recognition
4. Nature Language Process: NER(Named Entity Recognition) task, Sentiment classification, various tag task
5. Graph computing: Community detection and clustering
Basic knowledge you need：
1 probability theory
2 python or SQL basic
how to learn:
1 basic machine leaning course: get you familiar with the basics of machine learning such as feature engineering(feature synthesis and selection method) and Python
2 intermediate course: get you familiar with LR&GBDT theory and can independently design machine learning pipeline
3.advanced course: get you familiar with principles of various machine learning algorithms and can independently model new scenarios in the perspective of classification and regression(I will share the practice in commercial bank and leading internet campanies)

Subjects

  • Machine Learning Model deployment using flask and streamlit Beginner-Expert

  • AI & ML (Artificial intelligence & Machine Learning) Beginner-Expert

Experience

  • Senior Manager (Apr, 2022–Dec, 2023) at HONGKONG
    I、 online credit risk and anti-fraud project consulting in commercial bank
    Establishing online Small and micro enterprise credit risk full life cycle management system from scratch, including anti-fraud/application score card/credit limit pricing/industry differentiation strategy in pre-loan stage, behavioral model/credit approval in loan maintenance, LGD in delinquency management.
    Performance:
    1. Pushing the balance scale from 100 million to 17 billion+ by improving anti-fraud and credit risk interception capabilities.
    2. establishing data mart containing 3000+ features including history of repayment, financial statement analysis, personal credit analysis for credit risk identification.
    3. proposing automatic feature synthesis and evaluation method based on autocross and incremental evaluation algorithm which can reduce time consumption by 10 times.
  • AI Expert at Ant Group (Aug, 2017–Apr, 2022) at Hangzhou
    Constructing strategy search platform (SSP) and upgrading the pipeline of risk identification and interception in credit application process and apply it to a trillion-dollar balance credit products including loan product based on tax data/government procurement contract/Customs import and export data/e-commerce transactions data/bank statements, etc.
    Performance:
    1. traditional strategy development focus on the LIFT indicator, which can reflect the ability to distinguish between good samples and overdue samples. However, it does not consider the application approval rate. SSP employs Beam Search algorithm to find a set of strategies with the strongest ability to cover overdue samples under a given application approval rate constraint. this algorithm gives a standard for balancing business and risk departments, greatly improving the efficiency of strategy development goals and increasing the ability to recall overdue samples by 10%+.
    2. Bring graph computing into the field of credit risk and payment risk identification. Based on the transactions, social relationships, environment, etc., a heterogeneous graph with billions of vertices and tens of billions of edge relationships can be constructed. Then a fraud gang detection and identification algorithm based on community discovery is proposed. With the help of various correlations on the graph, this algorithm has raised the level of anti-fraud identification in the industry by 5%+ and reduces capital losses by more than 1 billion.
  • AI Expert at Alibaba (Jul, 2011–Jul, 2017) at Hangzhou
    Participate in the design of Alibaba Cloud algorithm components and be responsible for improving the accuracy of OCR, NLP, search ranking and other algorithms in business scenarios.
    Performance:
    1. traditional OCR algorithms adopts two-stage method of character segmentation + character recognition. Character segmentation does not consider context and it is difficult to solve the problem of Chinese left and right structure segmentation. Character recognition does not use language model decoding. An end-to-end semantic recognition solution is proposed base on FCN+CTC for arbitrary length OCR task. compared with the two-stage methods, Accuracy increased by 15%+ in the set of hard bad case.
    2. Based on CRF (LSTM+CRF, etc.) to realize the text structure. Defining various tag for business scenarios such as resume information extraction, NER, etc.
    3. employing Self-Attention mechanism between current search intent and member’s historical clicks and purchases to better understand e-commerce scenarios with search terms less than 10 words. This algorithm can improve the AUC of the DSSM by 1%+, which can significantly increase ECPM efficiency.

Education

  • automatic (Sep, 2009–Jul, 2011) from harbin–scored top 10%

Fee details

    HK$300/hour (US$38.56/hour)

    it's up to the project

Reviews

No reviews yet. Be the first one to review this tutor.