OLEDGuard ↗VR Chemistry LabAnti-Snore PlatformAI in LegalSimulation GamesYouTube Temporal SegmentationDynamic Movie SummarizationAI Music Generation+ ongoing
The systems behind the work — across Novartis, Optum, Nokia Bell Labs, Leadbook and NUS. Each diagram shows, step by step, how the system actually works.
Novartis · AI Innovation Centre2020 — present
Director, Data Science, AI & Computational Sciences — taking AI from prototype to validated production across discovery, R&D and enterprise; partnerships with NVIDIA, Microsoft and OpenAI.
Material approval · evidence
01/ 06material / claim received
Material / claim
needs approved evidence
Internal · lab reports
External · FDA · approvals
0
evidence docs
FDA label · pembro.95
Lab report · assay 7.88
Approval dossier.79
Internal SOP v4.71
Trial summary.63
document · page 12 · §3
TEXTp.12 · efficacy endpoint
TABLEp.7 · dose response
CHARTfig.3 · survival curve
IMGp.4 · assay image
SLIDEdeck · slide 9
chartfig.3 · survival.96
tabledose response.90
textefficacy endpoint.83
slidedeck · slide 9.74
imgassay image.66
Human reviewerin the loop
58%≥ 90% → auto-approve
01 · Novartis
Enterprise GenAI · Healthcare
Horizon MAP
An end-to-end multimodal platform that automates medical-claim and material validation. The hard part was the inputs: messy, unstructured lab reports with embedded charts and tables. Custom-built models parse those charts and tables into structured fields, a library of ~20 specialist models is orchestrated over them, and a unified evaluation layer with human-in-the-loop gates and full audit trails turns a multi-week manual review into a near-real-time, traceable workflow.
A sibling to Horizon MAP that turns the same governed foundation toward creation: an LLM system that generates marketing and communication materials — copy, layout and on-brand visuals — from a short brief and a set of brand and compliance rules, with review built in. The same orchestration and guardrails that validate documents are reused to produce them.
Brief + brand & compliance rules→Generate with an LLM→On-brand copy · layout · visuals
A first-of-its-kind agentic system for continuous, autonomous R&D, running as an iterative loop: ideation, then agents that build their own library of papers and tools end-to-end, exposed through a library agent over tools and MCP that both other agents and human researchers can query. Specialist personas — orchestration, planning, search, code, critique, writing — run across 10+ heterogeneous LLMs, with shared memory, human checkpoints, GPU/cloud orchestration, in-silico benchmarking and a comprehensive report after every cycle.
Ideate + build the library→Experiment + benchmark on GPU/cloud→Report · loop the cycle
An integrated antibody-discovery platform built around a reinforcement loop. Large-scale molecular-dynamics simulations are orchestrated across ~80 GPUs on DGX; their trajectories feed feature extraction that fine-tunes ML emulators of biophysical and developability properties — stability, aggregation, manufacturability. Predictions are benchmarked against wet-lab results (lab-in-the-loop), and the gap drives the next round. The emulators cut compute cost while preserving accuracy, alongside de novo design and binding-affinity optimisation.
MD simulation · 80 GPU→Trajectories → features→Fine-tune ML emulators→Benchmark vs lab → repeat
≈ 40%faster screening
~80GPU MD orchestration
2025Novartis Galaxy Award
Molecular dynamicsML emulatorsReinforcement loopGenerative de novo designAffinity optimisationNVIDIA DGX Cloud
Agentic RAG · evidence
knowledge base
multi-agent RAG
extract + verify
developability risk
VH/VLclinicalFDApatents
05 · Novartis
Agentic AI · Immunogenicity
Immunogenicity Evidence Extraction
A multi-agent RAG system that mines and validates immunogenicity evidence from VH/VL sequence-level signals all the way to clinical findings, FDA approvals and labels, patents and published studies. Extractor agents pull candidate evidence; critic agents challenge and cross-check it; a validation step enforces structured, dataset-level QA with traceable provenance — feeding downstream developability and risk assessment with evidence you can audit back to source.
Retrieve from sequences → labels → literature→Extract, then critic agents challenge→Validate with traceable provenance
AI-driven therapeutic-target identification that integrates multi-omics and high-dimensional biology with graph-based learning to surface and rank novel targets. A companion biomarker-discovery effort (imaging + proteomics) advances predictive signatures across neuroimmunology programmes, including progressive multiple sclerosis.
More from this eraPKS — drug-report extraction text · tables · chartsMicrosoft AI Empowerment 2021Responsible-AI governance EU AI Act
Optum · UnitedHealth2018 — 2020
Principal Data Scientist, AI Innovation (iLab) — production AI for affordability, payment integrity and clinical decision support inside a regulated US health plan.
Claims → spend driver
embed claims
temporal drift
across layers
flag driver
facilityprovidermember
07 · Optum
Payment Integrity · Custom Embeddings
Bi² — Behaviour Signal Intelligence
A system that transforms raw claim data into temporal signals using a custom-built embedding model, then tracks how those embeddings move through space and time across different layers of the healthcare system — facility, provider, member. Where an embedding drifts away from its neighbours, a previously invisible medical-spend driver surfaces — turning claims into prioritised, explainable affordability levers, with dynamic dashboards to explore them.
Embed claims with a custom model→Track drift across space · time · layers→Surface the spend driver
An engine that automates medical cost-saving ideation. It reads clinical and claims signal across the population and translates it into a ranked set of actionable, explainable affordability levers — surfacing where care can be delivered better and cheaper, and handing analysts a prioritised worklist instead of a blank page.
Clinical + claims signalCost-saving ideationRanking / prioritisationExplainable AI
Payment integrity · NLP + XAI
claim queue
NLP read
XAI explain
decision
✓ pay⚑ review
0 claims / hr
09 · Optum
Payment Integrity · NLP + XAI
AutoD — Automatic Payment Decision Helper
A high-throughput assistant for claim review and payment integrity. NLP reads each claim, an explainable-AI layer surfaces the reasons behind every call, and the system recommends pay-or-review under strict audit and compliance constraints — keeping a human in control while clearing the routine volume fast.
NLP reads the claim→XAI explains the call→Recommend · pay or review
NLP + XAIexplainable review
audit-gradecompliance constraints
HITLhuman stays in control
NLP / NLUExplainable AI (XAI)Payment integrityAudit & compliance
More from this eraUnsupervised fraud & anomaly detection claims-scaleProvider behavior monitoring self-supervisedDynamic dashboards D3 · Plotly
Nokia Bell Labs2016 — 2018
Lead Data Scientist, Innovation & AI — multimodal AI, large-scale media systems, and human-sensory interfaces (EEG / ECG / EMG / eye-tracking).
Text · image · video → one index
0indexed items0modalities
08 · Nokia Bell Labs
Multimodal AI · IEEE
Smart News Aggregation
The largest smart news-aggregation system built in the lab: it ingests news as text, image and video, classifies it with multimodal models, and organises everything under shared, machine-generated topic labels so the whole corpus becomes searchable and linkable across formats. The indexing core (ANNOTATE) was demonstrated at Mobile World Congress 2018 and published at IEEE Big Data.
Ingest text · image · video→Multimodal classification→Unified index + search
Work on AI-driven, passive human–computer interaction using EEG, ECG, EMG and eye-tracking. The flagship demo: a “mind-reader.” You think of an object and stand before a gaze-tracking screen; as images cycle, the model reads where your eyes settle and walks you down the ImageNet hierarchy — narrowing from broad categories to specifics — until confidence passes 80% and it returns its top-5 guesses. No clicks, no typing; intent inferred from gaze.
Read gaze on cycling images→Descend the ImageNet hierarchy→Confidence > 80% → top-5
A scalable pipeline that turns raw video into structured, multimodal object chunks. It ingests media at scale, segments each video, and runs fast deep-learning models to label every segment — finding the inner similarities and relations across a library so that video becomes searchable, linkable content rather than an opaque stream.
Video segmentationDeep learningMedia ingestionSimilarity / relationsSpark
Meaning · infer → compact → personalise
content
infer meaning
compact rep.
personalise
hier. LDAdisambig.transferper-user
11 · Nokia Bell Labs
NLP · Personalised Meaning
Smart Communication — Meaning Transformation
A system to represent, infer and communicate meaning for personalised content. Unsupervised hierarchical topic models — with experiments in deep generative networks — learn a compact representation of meaning that can be transferred and re-expressed per user, with evolving disambiguation of senses and generalised topic labelling across multimedia, so the same message adapts to each recipient.
Infer meaning · hierarchical topics→Compact, transferable representation→Personalised re-expression
More from this eraIndoor navigation & IoT context 2 patentsEEG / EMG passive interaction UX research
Leadbook2014 — 2016 · Singapore
Senior Data Scientist & R&D Lead — built one of Asia’s largest B2B intelligence platforms, end to end from data engineering to learned recommendation.
B2B intelligence graph · matched
0contacts0companies
10 · Leadbook
B2B Intelligence · 11 Patents
B2B Graph & Prospect Recommender
One of Asia’s largest B2B intelligence graphs — tens of millions of verified company and contact records merged from across the web — with a prospect recommender built on a patented Company–Product–Customer “deep relationship” model that learns which new prospects resemble a customer’s best existing ones. The engineering ran from distributed crawling and entity-matching to real-time lookup.
Crawl & merge millions of records→Model company–product–customer ties→Recommend the right prospects
44M+verified contacts
11.5Mcompanies
11Singapore patents
Deep Relationship ModelSpark · HadoopElasticsearch (custom plugin)Go proxyReact extensionTorch
More from this eraReal-time lookup proxy GoContact-lookup browser extension React.jsPersona-scoring plugin ElasticsearchFeature classifier TorchEmail-validation service
NUS & Singapore2008 — 2014
PhD & Researcher, School of Computing (A*STAR SINGA) — NLP, topic models and privacy-preserving data systems, alongside ventures and AI-for-good prototypes.
Discussions → aspect / action topics
pricing · complaint.92
feature · request.85
support · praise.77
delivery · issue.69
UX · suggestion.61
11 · NUS (PhD)
NLP · AAAI
Aspect–Action Discussion Graph
Doctoral work that takes millions of flat user comments and posts and turns them into a temporal aspect–action graph: a joint aspect–action topic model infers what people are talking about and what they intend to do — without labelled data — and arranges it into a structured, time-aware hierarchy of who said what about which aspect, when. Published at AAAI; the foundation of a self-supervised discussion-analysis and prediction system, supervised by Prof. Chua Tat-Seng.
Millions of flat posts→Joint aspect–action topic model→Temporal aspect/action graph
A wearable smart ring that fuses on-device AI vision with ultrasonic / sonar ranging to perceive obstacles and open space, then guides blind and low-vision users with intuitive, real-time directional feedback. Designed and prototyped in Singapore.
Point a phone at a supermarket shelf and a vision-and-health model recognises each product and paints a personalised health “hue” over it — guiding shoppers toward better choices for their own profile. It runs on the first structured database of Singapore food labels, which we built and processed to train the model.
Recognise each product→Score for your profile→Paint a health hue
1ststructured SG label DB
Personalper-profile scoring
Phonelive camera overlay
VisionOCRNutrition modellingCustom datasetMobile / AR overlay
Location → crowd tasks
where you are
contextual topics
match tasks
crowd shoppers
14 · Singapore
Venture · Startup Weekend
MysteryShopper
A crowd-sourced, location-based mystery-shopping platform that matches tasks to the right people and places using contextual topic models — turning everyday shoppers into a distributed sensing network for retail insight. Winner, Startup Weekend Singapore.
Contextual topic models→Match task · person · place→Distributed sensing
WinnerStartup Weekend SG
Geolocation-aware matching
Crowddistributed network
Topic modelsGeo-matchingCrowdsourcingMobile
More from this eraClinical anonymization — PASS DASFAA 2010Maritime geolocation graph · topic modelsVisual model for the blind GTC 2015
Independent & open projects
Things I build on my own — products, prototypes and systems exploring AI in everyday life, the law, training and the body.
intake
review
strategy
deposition
radar
privilege ✓
a matter → six modules → defensible output · audit trail
AI · Legal Platform
The AI Cockpit
A comprehensive AI platform for legal teams — six connected modules that carry a matter from intake to defensible output, with cited reasoning and an audit trail running through all of it.
Smart Inbox & Matter TriageReads, routes and opens matters from the inbox
Document ReviewReviews documents and answers with citations
Strategy WorkbenchBuilds and pressure-tests case strategy
Deposition ArchitectPlans depositions, reconstructs the evidence timeline
Clients & Regulatory RadarTracks clients and shifting regulation
Privilege SentinelGuards privilege and keeps the audit trail
A physics-based, pixel-level OLED guardian that protects against burn-in — and keeps working inside games by automatically detecting static HUD elements and treating them in real time.
A VR chemistry-lab training simulation, prototyped with a custom liquid-interaction engine and a library of lab protocols — so learners can pour, mix and run procedures with believable fluids in a safe virtual lab.
A personal operating system of AI agents that helps you run your life — capturing, planning and offloading the mental tasks you'd otherwise juggle, so more of the day's overhead runs itself.
A system that helps you get 1% better across the parts of life you care about — small, AI-guided adjustments, tracked over time, compounding into real change.
Habit trackingAI nudgesCompounding goals
0processes
directed steps · follow · fork · improve
Knowledge · Processes
Global Process System
A “Wikipedia of processes” — one place gathering the steps for anything with a procedure: applying for a job, onboarding at a company, a visa application, even becoming an Olympic medallist — structured so you can follow, fork and improve them.
Structured processesVersioningCommunitySearch
Recognition
13+
Patents filed
2 US (behaviour-signal & data-perspective, Optum), 11 Singapore (B2B prospecting, Leadbook), plus 2 at Nokia Bell Labs (IoT context & indoor navigation).
5+
Peer-reviewed publications
Venues include AAAI, DASFAA and IEEE Big Data — spanning topic modelling, multimodal indexing and privacy-preserving clinical NLP.
★
Awards & honours
Novartis Galaxy Team Award (2025) for the Biologics AI Platform, Star Award for Leadership (2022), and Winner, Startup Weekend Singapore.
