Zhixiang Feng

Current ML research work at Revax, focused on improving the neural reranker behind a UK tax retrieval augmented generation workflow.

• Improving retrieved evidence quality before answer generation
• Analysing hard tax queries and reranker failure cases
• Turning evaluation findings into practical reranker changes

AI for healthcare research on generating synthetic multi-label epilepsy datasets from a small set of clinical letters.

• Investigating synthetic labels for scarce real-world medical text
• Designing checks for label diversity, accuracy, and reliability
• Targeting more than 15,000 synthetic cases from fewer than 100 real samples

Research on how adversarial attacks affect neural and non-neural retrieval pipelines in search systems.

• Implemented recent adversarial attack algorithms targeting ranking components
• Simulated search pipelines that combine neural and non-neural retrieval
• Analysed how attacks changed document order, relevance scores, and search behaviour

Chrome extension built in a University of Glasgow team project with Amazon.com, Inc. as client, combining a page indexer, markdown notebook, and to-do list.

Project page: Hawk - Page Indexer on the Chrome Web Store

• Implemented local page indexing and search over selected web content
• Used BM25F with MiniSearch fallback for ranked document retrieval
• Supported JSON import and export for indexed page backups
• Published as Hawk - Page Indexer on the Chrome Web Store
• Presented a final product recognised by software engineers at Amazon Edinburgh

First place hackathon entry for MS Tanks, using a TCP API, an epsilon-greedy strategy, and a subsumption architecture.

• Built an autonomous tank agent for the Do You Have the GUTS Hackathon 2024
• Balanced exploration and exploitation with an epsilon-greedy policy
• Prioritised survival and resource search through subsumption based behaviours

Question answering project using WikiQA, set similarity methods, term frequency vectors, and BERT based context vectors.

• Compared overlap coefficient, Sorensen-Dice, and Jaccard similarity
• Evaluated BERT based answer selection with accuracy, precision, recall, and F1
• Analysed model stability across different question and answer scenarios

Mobile event registration app for a non-profit organisation supporting children with autism and their families.

• Built event sign-up flows for parents and children
• Added organiser tools for venue capacity and booking management
• Connected mobile UI work with backend functionality

Recipe sharing web app where users could publish recipe blogs, rate posts, save favourites, and explore trends.

• Worked on requirements analysis, UI design, technology selection, testing, and deployment
• Built posting, editing, rating, and favourites functionality
• Used the project to practise full stack delivery from schema to interface