Experience

Selected work

Defence Science and Technology AgencyJanuary 2021 – Present · Agentic AI lead since 2025

Agentic AI Product Lead, Cyber Defence

  • ProblemDays of manual hunt planning per threat report
  • My roleProduct direction, architecture, and team delivery
  • OutcomePlanning reduced from days to hours

Spearheaded the design and implementation of a LangGraph-based agentic threat-hunting solution that turns threat reports into hunts. Orchestrated narrow ReAct and external A2A agents with defined tool access, guardrails, and analyst review points. Led the context engineering and evaluation needed for reliable performance on a local LLM in an on-premises environment while reasoning over large volumes of telemetry. Validated the solution through red-team evaluation.

  • Multi-agent systems
  • LangGraph
  • LLM evaluation
  • Human-in-the-loop

Click the chartTap the chart to see it in action

Agentic Threat Hunting|From a ransomware report to one lead1Purpose-built agentsA local model does best with one job, soeach narrow agent gets only the toolsand context that job needs.2Equipped to adaptNo results can mean a bad query, not asafe network. So the agent rewrites it,checks the fix and searches again.3Surface to the analystAgents flag what looks suspicious, withthe evidence attached. Only a humananalyst makes the call.SupervisorHypothesisRansomware reportReportingHunch: attackers hide in scheduled tasksHunch: attackers hide in scheduled tasksEvent logs0foundfound481216202428321 unusual>process="schtasks.exe">event="task created"Empty doesn't mean safeFields and syntax checked1 leadScheduled TaskATT&CK T1053.005A known attacker trickAnalyst decidesDetectionAgentic Threat HuntingFrom a ransomware report to one leadSupervisorHypothesisDetectionRansomware reportReportingHunch: attackers hide in scheduled tasksHunch: attackers hide in scheduled tasksEvent logs0foundfound481216202428321 unusual>process="schtasks.exe">event="task created"Empty doesn't mean safeFields and syntax checkedDetection1Purpose-built agentsA local model does best with one job, soeach narrow agent gets only the toolsand context that job needs.2Equipped to adaptNo results can mean a bad query, not asafe network. So the agent rewrites it,checks the fix and searches again.3Surface to the analystAgents flag what looks suspicious, withthe evidence attached. Only a humananalyst makes the call.1 leadScheduled TaskATT&CK T1053.005A known attacker trickAnalyst decides

Also at DSTA · January 2021 – Present

Cyber Detection Analytics

Developed custom analytics to detect anomalous activity in AWS and Microsoft 365, and phishing domains impersonating MINDEF websites. Worked with stakeholders on use cases and methods, then rolled out what worked.

  • ProblemIdentifying deviations in high-volume cloud activity, and sites impersonating MINDEF
  • My roleAnalytics roadmap, technical direction, and delivery
  • OutcomeNew detection capabilities in daily operational use
  • AWS
  • Microsoft 365
  • Behavioural baselining
  • Data science
  • Python

HoverTap a chart to see it catch something

Phishing domains

Copies of real sites are flagged.

Statistical baseline

Points outside the normal range are flagged.

Graph analysis

Links never seen before are flagged.

Early career Internships

ShopeeData Science InternApril – November 2020 · 8 months
Singapore

Multilingual Product NLP

Built models that predicted product attributes across five Southeast Asian languages.

  • NLP
  • Multilingual ML

Hover the chartTap the chart to see it in action

1/9 Rp189.000-30% Sepatu Lari Pria 4.9Terjual 2,3RB Gratis Ongkir Spesifikasi Produk SepatuLariPriaIDSepatuIDLariIDPriaWordsBreaks the titleinto tokens. IDModelPredicts theattributes. CategoryShoesTypeRunningGenderMenActivitySportAttributesFills in theproduct specs. 1/9 Rp189.000-30% Sepatu Lari Pria 4.9Terjual 2,3RB Gratis Ongkir Spesifikasi Produk Sepatu Lari Pria IDSepatu IDLari IDPria WordsBreaks the titleinto tokens. ID ModelPredicts theattributes. CategoryShoes TypeRunning GenderMen ActivitySport AttributesFills in theproduct specs.
The LeagueData Science InternAugust 2019 – April 2020 · 9 months
San Francisco, USA

Fraud & Risk Detection

Built a risk model that cut manual profile review by 70%, plus alerts for 20+ business metrics.

  • Fraud modelling
  • Product analytics

Hover a chartTap a chart to see it in action

5:00 THEleague James M.45 • Deployed • 6’ 1” EducationProfessionInterestsGroupsPhD, EngineeringUS ArmyGod • Family • Church • Honest& Caring • TravelCatholic Singles • WidowersSupport • More Risky profile SignalsModelRisk Photo, bio andother clues.A classifierweighs them.Risky profiles goto manual review. 5:00 THEleague James M.45 • Deployed • 6’ 1” EducationProfessionInterestsGroupsPhD, EngineeringUS ArmyGod • Family • Church • Honest& Caring • TravelCatholic Singles • WidowersSupport • More Risky profile SignalsModelRisk Photo, bio andother clues.A classifierweighs them.Risky profiles goto manual review.
AB5:00THEleagueMost compatibleJess M.41 • Presidio • 5’ 6”EducationProfessionUCLAProduct Designer5:00THEleagueMost compatibleJess M.41 • Presidio • 5’ 6”EducationProfessionUCLAProduct DesignerMatch rate by designHow likelyA · BorderB · BadgeMatch rateRandomiseMeasureEach user seesA or B at random.B's match rate is higher,and it's significant. AB5:00THEleagueMost compatibleJess M.41 • Presidio • 5’ 6”EducationProfessionUCLAProduct Designer5:00THEleagueMost compatibleJess M.41 • Presidio • 5’ 6”EducationProfessionUCLAProduct Designer RandomiseEach usersees A or Bat random. Match rate by design How likely A · BorderB · Badge Match rate MeasureB's match rateis higher, andit's significant.
NUHSData Science InternDecember 2018 – July 2019 · 8 months
Singapore

Clinical NLP & Readmission Prediction

Predicted 30-day readmissions, automated clinical text extraction, and created realistic synthetic data.

  • Clinical NLP
  • Synthetic data
Earlier projectsHackathons and showcases
BrainHack 2022 virtual command centre
June 2022

BrainHack CDDC

Helped organise and plan a cybersecurity capture-the-flag event as part of its management committee.

View BrainHack
Devil's Whisper technical showcase title card
June 2021

Devil's Whisper

A technical showcase explaining adversarial audio attacks on smartphones and smart-home devices.

View showcase
Hack4Climate team after winning the data-science category
March 2019

Hack4Climate

First place in the Yale-NUS Datathon data-science category for an environmental responsibility story.

View project