Juncen Ye Email & Phone Number

Paris, Île-De-France, France

Paris, Île-De-France, France

- Developed a Q&A prompt using LangChain to extract medical metadata from clinic websites by breaking content into chunks, refining the question in the prompt, and segmenting data to highlight investment potential.- Created interactive dashboards with customized Shape Maps in Power BI for data visualization.- Detected vehicle models from images using EfficientNetV2 (achieving 86% accuracy) and enhanced detection with YOLOv7 and TrOCR.- Transformed L’Oréal’s unstructured data into structured formats by leveraging Azure Data Lake for scalable storage and real-time processing, and Google Cloud Platform (GCP) for data transformation pipelines, including the integration of BigQuery for advanced data analytics and Cloud Storage for optimized workflows.Code: Python, SQL, Git, OpenAI API, Google Map API.Tools: LangChain, Azure Data Lake, GCP (BigQuery, Cloud Storage), Power BI.

Paris, Île-De-France, France

- Created a 500k metadata database based on the Reverso grammar correction database (original sentence/corrected sentence), containing 50 parameters such as grammatical composition, keywords, and sentence subjects.- Developed grammar correction datasets (categorized by grammar error type) using online datasets such as CoNLL, FCE, and BEA, as well as authorized Reverso user data.- Fine-tuned models such as Gramformer and GECToR using Docker containers, training them on grammar correction databases collected and corrected from real users.- Optimized the model comparison tool MTcompare, applying various evaluation metrics such as the GLUE score for performance analysis.- Scraped a comprehensive pinyin (Chinese pronunciation) database from Wiktionary and merged it with the existing Reverso pinyin database to correct errors and include new pinyins for words and phrases.Code: Python, SQL, Git, Docker.Tools: Tensorflow, Huggingface, Gramformer, GECToR, MTcompare.

Paris, Île-De-France, France

- Scraped 8 million tweets generated by RATP passengers on Twitter using Twint, Tweepy, and Twarc API tools, and established a comprehensive real-time metadata database containing 46 features (including tweet ID, conversation ID, tweet language, etc.).- Applied a BERT model for topic modeling to classify tweets with similar topics and utilized TextBlob for sentiment analysis.- Processed metadata using Pandas, assisting the team in extracting informative data from the database for relevant data insight presentations. Created visualizations using Seaborn and pyecharts for enhanced data representation. Code: Python, Twitter APIs.Tools: Twint, Tweepy, Twarc, Transformer, TextBlob, Pandas, Seaborn, pyecharts.

- Investigated the performance of TrouSPI-net(a network that is designed for pedestrian action prediction) on two gesture sequence classification tasks: SHREC and JHMDB.- Analysed the structure of TrouSPI-net including Convolutional Block Attention Module (CBAM) and U-GRU blocks.- Applied Knowledge Distillation to train a 'student' model that decreases the network size for more practical use.- Obtained 95.3% and 90.4% recognition accuracy with only 40k parameters on SHREC 14 and SHREC 28 data set and compared with other methods such as CNN+LSTM and DD-Net.

- Responsible for front-end functional programming (Javascript) with React Native.- Created the ER graph and established the backend database with SQL correspondingly.- Achieved functions such as searching, proposing and checking a trip.- Optimised the front-end UI/UX with the help of expo packages.

- Applied web scraping("request" and "beautiful soup") to get the online news information.- Processed data cleaning such as lowering the capital letters, removing stop words and putting the word bags into matrices by applying "pandas".- Divided news into different categories, in each category we divided news into different topics, finally in each topic we divided it into different opinions by applying LDA.- Captured the polarity and subjectivity of different options by applying "TextBlob".- Visualised the results with "pyechart" and "seaborn".

- Applied EULA and QUA to simulate the information of hand rotation.- Introduced Attention Mechanism in SLR and created an end-to-end SLR framework in order to train the model effeciently.- Applied a multi-channel CNN to enhance the model's ability to deal with different user data.- Attaned 1.3% average error rate per sentence and 0.3 second real-time recognition speed.

Ingénieur Civil, Deep Learning/Data Science

- Deep learning for image analysis - Artificial Intelligence - VR 3D museum - Artificial Intelligence - Mobapp(simulate mobile application.

Bachelor'S Degree, Computer Science

Activities and Societies: Head of Design Department of Student Union: Design posters, leaflets, banners and other promotional products /.

Exchange, Computer And Systems Sciences

Exchange to the Department of Computer and Systems Sciences with CSC(China Scholarship Council) overseas study scholarship. - Passed all.