I turn complex data into clear decisions

Situation

A local Malaysian bank's auto-financing loan applications were submitted by partner officers via email with attached documents (NRIC, payslips, bank statements). Bank officers had to manually review each email, open every attachment, and key customer details into the loan origination system — a slow, error-prone process.

Approach

• Participated in client requirement gathering to scope the AI email processing solution
• Assisted in designing the Azure architecture (Document Intelligence, Logic Apps, Blob Storage, Functions)
• Contributed to building document classification, OCR/NLP data extraction, and a cross-validation engine
• Built audit trail reporting with extracted fields, mismatches, and confidence scores

Impact

Delivered a working PoC that automates document classification, data extraction, and cross-validation of customer details — enabling loan officers to focus only on exception handling. Designed to reduce processing time, manual errors, and improve approval consistency.

Situation

Clinical variant interpreters were manually reviewing multiple public sources (PubMed, ClinVar, dbSNP, FDA, clinical trials) to identify variant evidence. Existing tools provided inconsistent literature synthesis and required costly full-suite subscriptions, wasting time and budget.

Approach

• Contributed to building the report verification module on Google Cloud Platform (Cloud Run, BigQuery, Vertex AI)
• Developed tier validation, variant interpretation checks, and mutation detection flags with human-in-the-loop indicators
• Implemented structured logging with unique identifiers for traceability across concurrent modules
• Helped resolve report generation bottlenecks through worker provisioning

Impact

Reduced clinician verification time through automated literature summarization with clear flags for expert review. Delivered client-specific customized reports. Achieved ~90% cost reduction compared to traditional interpreter tool subscriptions.

Situation

SkyWorld Development Berhad, a leading Malaysian developer, needed a scalable way to handle multilingual customer queries and extract structured sentiment data to inform strategy. A university-industry collaboration.

Approach

• Ran stakeholder analysis to define chatbot scope and sentiment framework
• Engineered hierarchical sentiment models aligned with business classification tiers
• Designed modular FastAPI + MultiLLM architecture for multilingual support (Malay, Chinese, English)
• Trained models with Hugging Face pipelines, meeting accuracy KPIs

Impact

Delivered a working multilingual prototype with integrated sentiment analytics. Modular architecture ensured the client could maintain and extend it independently after handover.