Benjamin Chu Min Xian

• Visual Pillbox - Completed a Gen-AI proof of concept in collaboration with Khoo Teck Puat Hospital, as part of the AI Trailblazer initiative, leveraging Google’s Vertex AI and Retrieval Augmented Generation (RAG). I developed a system that processes medication images—whether single or multiple uploads—using Google Vision OCR to generate patient-centric summaries that include storage guidelines, administration instructions, and common side effects. It also produces a visual schedule that arranges medication doses into morning, afternoon, and evening slots throughout the week. The system accesses a vector database of medication leaflets for targeted context retrieval and utilises Vertex AI’s data source grounding to prevent hallucinations. With LLM chaining via LangChain’s MapReduce, it condenses large contexts into clear, concise summaries. This solution automates visual handouts for pharmacists, significantly enhancing patient education and communication.
• HealthKaki - a 14-week proof of concept, was developed in collaboration with the Ministry of Health's Public Health Division, Synapxe, and Temus to empower residents in managing their health through personalised recommendations. The PoC is completed, where my key role is to develop an agentic workflow using a stateful, graph-based mechanism within the LangGraph framework to dynamically generate and refine exercise plans. I designed the components to be modular and adaptable, employing state management for real-time updates, condition-specific decisions, and integration of Tier 1 health actions from the Ministry of Health and Health Promotion Board guidelines. Workflows offered personalised exercise recommendations by merging contexts from structured knowledge bases (e.g., HPB MoveIt series) and open-source data, enriched with metadata for targeted muscle groups. I integrated AWS Bedrock's Claude 3.5 LLM to create tailored exercise plans that considered user preferences, activity history, and health conditions, ensuring safety and customisation. Furthermore, I utilised AWS DynamoDB to store user settings, allowing personalised exercise plans to evolve based on previous configurations.
• C3 Graph - Completed a POC exploration with TTSH, focusing on optimising patient flow at triage through graph analytics to enhance bed utilisation and reduce congestion. Patient movement was represented as a graph network to visualise admission flow and monitor multiple transfers between wards until discharge. Graph clustering and centrality analysis were utilised to identify daily and hourly hotspots, highlighting wards with the highest incoming admissions. Predictive models leveraging graph-derived features were developed to forecast bed type requirements upon admission, providing a 1.5-hour lead time for proactive resource management. This approach facilitated earlier discharge planning and minimised supply-demand mismatches, improving overall patient flow efficiency.
• Patient Journey - Conducted exploratory work for HSA to leverage graph analytics to enhance the analysis of patient journeys. Developed patient-centric views with interactive widgets to summarise patient pathways, capturing visits, diagnoses, and prescriptions over time. Created sub-queries with filter parameters to generate detailed visualisations, including medication timelines, diagrams of diagnosis-medication interactions, and lists of prescriptions. Built a dashboard for cohort-level analysis, segmenting patients into groups and identifying trends in disease and treatment patterns using graph algorithms and traversals. The dashboard provides insights into diseases and medications, allowing users to quickly identify effective treatments and assess the efficacy of medical prescriptions and procedures after a diagnosis.

POV Scoping and Technical Implementation with TigerGraph
Led the scoping and solutioning of Proof-of-Value (POV) projects, defining technical implementations using TigerGraph functionalities to address client use cases. Assessed client data points to design graph schemas and execute data mapping, transforming raw data into TigerGraph's native representation for ingestion. Collaborated with sales teams during customer calls to understand problem contexts and draft detailed scopes, success criteria, and deliverables for POV plans. Developed graph logic and queries tailored to client-specific use cases, including path traversals and graph algorithms.

Key POV Projects:

• Singtel Network Data Connector (NDC): Enabled correlation of poor-performing cell towers with users connected to them at specific locations and times, allowing detailed drill-downs into network performance issues. Implemented use cases demonstrating the ability to identify low-experience cells and determine the affected users, pinpointing the number of impacted users and their experience scores at specified hours.
• GIC: Conducted deep link analysis to uncover hidden relationships between individuals and companies across multiple degrees of connection. Demonstrated capabilities to trace connections between GIC personnel through activities, meetings, and shared links. Implemented recommendation systems using graph algorithms to suggest meeting notes relevant to an individual’s network or recommend additional participants for meetings based on past interactions and similarities.

• Project IRIS: Designed and implemented a framework for risk propagation to identify and score potential suspicious actors in graph networks, utilising centrality measures and risk propagation mechanisms. Scored entities and subgraphs based on network density and structure, creating an anomaly detection module to flag unusual graph patterns. Developed a scalable graph analytics SDK, architecting a graph processing platform integrated with AWS Neptune as a data lake, along with custom adapters for AllegroGraph and Neo4J, and NetworkX for scoring pipelines.
• Sherlock: Developed a hybrid, unsupervised approach to topic modelling that combines customised LDA with Gibbs sampling, graph centrality, and clustering. Applied this generalised method to financial equity reports, earnings transcripts, and Reuters News, revealing actionable topics for downstream applications.
• Infura (Infrastructure 360°): Developed a robust methodology to integrate various infrastructure project datasets—financing, deal, and ESG data—using a graph-based approach. This solution addressed the investment gap in the infrastructure sector, where private equity capital struggles due to issues with data transparency. I have developed an implicit linking method that combines NLP and Refinitiv PermID technology, covering over 60,000 projects, 24,000 entities, 30,000 connected deals, and more than four million relationships across six linked databases, accessible via a RESTful API. Data exploration and visualisation were conducted in StarDog Studio, with RDF graphs exported to AWS Neptune for efficient querying and integration. A study report for the Global Infrastructure team highlighted the limitations of the previous SQL-based approach, particularly its rigid schemas and interdependencies. The graph-based solution enabled faster insights and allowed for detailed analysis of projects by type, deals, transactions, and related dependencies. By linking unstructured data—such as comments and text—through Named Entity Recognition (NER) and associating entities with unique PermIDs, the knowledge graph significantly enhances data access and transparency, thereby improving infrastructure data visibility from existing datasets.

• Toggle TV Graph: Designed and developed a graph-based search engine for Toggle, organising media content such as series, episodes, and clips into a graph structure. Implemented relevance, phonetic, and semantic search algorithms to enhance user search experiences.
• Exploration Project with DBS: Built a linguistic analyser to classify verbatim feedback (short comments indicating praise or complaints) related to specific branches and services. Leveraged machine learning models to categorise verbatims into themes and applied a hybrid approach combining linguistic patterns and ML for sentiment classification.
• Auto News Tagger for Channel NewsAsia: Developed an enhanced auto news tagger to analyse news articles and generate relevant topic-based keywords and named entities. Utilised DBpedia for name resolution and topic identification, and Internet Advertising Bureau (IAB) standards for generating relevant keyword topics.
• Project Greenlight: Designed and built a movie analyser, "Greenlight," to evaluate movie synopses. Integrated predictive models, natural language processing (NLP), and Dynamic Time Warping (DTW) to analyse sentiment shape sequences within storylines. Classified synopses into Blake Snyder’s story archetypes and IMDb categories. Employed Shapley attribution values to assess the influence of specific story archetypes on a movie's success, providing actionable insights into narrative-driven factors.

• Marketplace AI: Designed, tested, and refined a predictive model for Marketplace AI on 11street, one of Malaysia’s largest e-commerce platforms, a joint venture with Celcom Axiata Berhad and SK Planet. The model scanned product listings, including labels, descriptions, and OCR-extracted text, to identify and classify illegal or banned substances, counterfeit products, e-cigarettes, and other prohibited items. Each product was scored against designated risk categories.
• OneSentry AI: Developed and optimised a classifier model for OneSentry, a solution to monitor merchants for high-risk behaviours during onboarding and operations. Implemented a web heuristic module to assess merchant websites and social media platforms for traits indicative of risk, such as non-compliance with regulations on pharmaceuticals, gambling, and deceptive practices. Utilised sentiment analysis to detect negative customer sentiment, a key indicator of chargeback risks, enabling proactive compliance monitoring.
• 11street POC: Leveraged Neo4J to create a methodology for detecting coupon fraud by extracting network insights.
  Developed a spam detection system for the marketplace using Natural Language Processing (NLP) to extract linguistic patterns. Applied topic modelling and predictive algorithms to identify outliers and potential fraud.
• FraudWall AI: Led the design of FraudWall AI, transforming transactional records into graph structures. Utilised graph analytics to identify anomalous patterns, enabling detection of suspicious transactions and strengthening fraud prevention rules.

• Text-to-Graph Platform Development: Designed, and developed a Natural Language Processing (NLP) platform for knowledge analytics and extraction from unstructured texts in English, Malay, and Mandarin. Implemented web-service-enabled components to identify entities and relationships, transforming text into knowledge graphs stored in AllegroGraph for querying, insights generation, and visualisation on a dashboard. Benchmarked and published research papers evaluating the platform's components.
• Algorithm Design & Complexity Analysis: Analysed module complexities and designed algorithms to optimise language processing workflows across multiple languages.
  Authored requirement and design documentation, conducted functional and performance testing using tools such as JMeter and SoapUI, and guided testers in preparing comprehensive test plans.
• Leadership & Team Management: Led a multidisciplinary team in researching, developing, and testing individual modules of the NLP platform.
  Oversaw benchmarking and evaluation processes that led to research publications and patent filings for the platform's innovative components.
• Academic & Research Collaborations: Directed collaboration with Universiti Malaya on an E-Science-funded project to develop a Malay Text Understanding system for processing unstructured Malay texts.
  Coordinated partnerships with local and international institutions, including Universiti Malaya (UM), Multimedia University (MMU), the German National Library of Economics (ZBW), and the Food and Agriculture Organisation (FAO), fostering knowledge exchange and research advancements.
  

• Business Analysis: Functioned as a business analyst to evaluate client requirements, translating business needs and process flows into detailed requirement and design documents.
  Delivered documentation crucial for developers to build key modules for the FAME wealth management platform.
• System Development: Contributed to developing new features and enhancements on the FAME platform, including writing SQL stored procedures for transactional modules.
  Customised and improved the platform interface, resolving issues and ensuring seamless functionality.