Case studies
Recent engagements, and what they returned
Situation
CBRE needed a faster way to extract tabular data from real estate listings. Staff were manually checking and collating data across multiple PDF pages, which was slow and error-prone.
Approach
Built a Python pipeline that routes each document by structure: Camelot for tables with defined borders, the LlamaParse API for irregular or unstructured tables, with a header detector and Excel output compatible with CBRE’s existing database.
Result
25% time saved and 95% extraction accuracy, with robustness demonstrated across the range of listing formats in use.
Situation
The school wanted to understand what drives student demand across 20+ courses to support long-term curriculum planning.
Approach
Combined clustering to group similar modules and cohorts, panel regression to capture multi-semester dynamics, and time-series analysis to forecast enrolment.
Result
Identified the key drivers of module demand and cohort preferences, and delivered recommendations that informed curriculum design and enrolment strategy.
Situation
Inventory monitoring relied on manual tracking, consuming staff hours on tasks that could be automated.
Approach
Built a Tableau trend-simulation workflow to model inventory movements and surface forecast signals automatically.
Result
More accurate inventory forecasts and substantially less time spent on manual monitoring, freeing hours previously lost to automatable tasks.
Situation
The Environmental Services Office needed a clearer picture of portering demand across the hospital to plan daily staffing.
Approach
Analysed portering service data to identify demand patterns by time and location, then modelled manpower allocation against those patterns.
Result
Operational recommendations on daily portering schedules and manpower allocation, adopted for planning by the Environmental Services Office.
Situation
Manual review of unstructured US SME bank statements was time-intensive and error-prone, creating bottlenecks in underwriting and risk assessment.
Approach
Built and deployed a transaction-enrichment pipeline using AI classification, converting unstructured statements into structured categories with “Reason” tags.
Result
Review time fell from 25–35 to 5–10 minutes per statement, with 90–95% automated accuracy. Validated by the CrediLinq.AI CTO.
Other notable projects

American Express
Tableau dashboard presenting profit and loss in a comprehensive, interactive format.

Singapore Police Force
Survey analytics consultation for the operations department.

Lex Quanta
Law analytics: predicting judgments from historical cases.

Datorama
Automated data scraping in R with an NLP prototype to cut extraction time for a risk-intelligence platform.

NUS School of Computing
Academic research with professors from the Business Analytics Centre.




