Ni LikunData Science · Singapore Download CV
NewSubmitted to the LIFT Challenge 2026

MSc Data Science for Sustainability · NUS

Ni Likun

I turn human-mobility and geospatial data into evidence-grounded AI for smarter cities.

Research Assistant at the NUS Institute of Data Science, working on spatiotemporal data mining and smart-city simulation. Previously at LocationMind (Tokyo) and The University of Tokyo, building geospatial ETL and an early LLM-agent prototype for mobility prediction.

  • Human mobility
  • Geospatial intelligence
  • LLM agents
  • Data analytics
t−5t next location? candidatecandidate
Illustration Next-location prediction from a stay trajectory, with routine history and candidate-location evidence.

News

Latest updates

Pinned Job search

Looking for full-time roles in Singapore

Data analytics, applied AI and smart-city / geospatial roles — where mobility, spatial and text data meet LLM-agent workflows. Currently a Research Assistant at the NUS Institute of Data Science; MSc expected Jul 2027.

  • Data Analyst / Data Scientist
  • Applied AI · AI Agent Product
  • Smart City & Mobility Analytics
  • Geospatial Intelligence
Download CV Get in touch
UrbanPulse index map of Singapore: HDB blocks coloured from less to more heat-resilient 10,749 HDB blocks scored
Case competition

Launching Investment for Future Transformation (LIFT) Challenge 2026

Our entry: UrbanPulse

Submitted our blended-finance proposal to the LIFT Challenge run by the Ackerman Center for Excellence in Sustainability, UNC Kenan-Flagler. UrbanPulse is a block-level heat-resilience and inclusion index for all 10,749 HDB blocks in Singapore, built from Landsat, Sentinel-2, OpenStreetMap and 16 data.gov.sg datasets. It adds a community-data loop and a green-finance model linking verified outcomes to sustainability-linked loans.

  • 24 pts largest inclusion gap hidden inside a town average
  • −0.48 °C per +10 pp tree canopy within 200 m
  • −41% 30-year greenery cost on a GLS case site
Website GitHub LIFT Challenge ↗

At a glance

Results I have contributed to

+11.4pp

Top-1 next-location accuracy

60.05% → 71.42% on BW · LLM-agent mobility study (team result)

−36%

Mean location error

3.44 km → 2.20 km · same study, final system

1.8M

Records collected & structured

~270k races · 100k+ node pedigree network, built independently

ACM ’26

Peer-reviewed publication

MACA-PA multi-agent clinical assistant · HIDA 2026

What I bring

Three strengths, each backed by real work

Geospatial & mobility data

Turning raw GPS, POI and urban layers into clean, analysis-ready features.

  • Built geospatial ETL inputs at LocationMind, Tokyo
  • Joined urban, socioeconomic & infrastructure layers into POI location features
  • Trajectory preprocessing across three mobility datasets

Evidence-grounded LLM agents

Designing agents that reason over retrieved evidence — and evaluating whether they actually help.

  • Built the first AgentMob prototype (API / RAG tool calls)
  • Benchmarked large multimodal models on satellite-image spatial reasoning
  • Co-authored a multi-agent, local-RAG clinical assistant (ACM HIDA)

From analysis to decisions

Statistics training plus clear writing — comparisons, diagrams and reports people can act on.

  • BSc Statistics, Outstanding Thesis (network community detection)
  • Compared mobility & spatial-analytics methods for product reviews
  • Publication-ready workflow diagrams and manuscript editing

Featured research

Evidence-grounded mobility prediction with an LLM-driven agent

Mar – Aug 2025 · Information Technology Center, The University of Tokyo · arXiv:2606.05130 ↗

The question

Sequence models capture historical movement well but miss why people travel. Can an LLM agent add semantic travel intent and geographic context to next-location prediction — without paying LLM cost on every case?

The approach

Routine cases take a cheap history-based fast path; ambiguous ones go to tool-assisted reasoning that gathers evidence — recent context, historical behaviour, stay/move likelihood and candidate-location profiles — across three mobility datasets.

My role

  • Scoped the research question and designed the initial workflow.
  • Built and validated the first Python pipeline: trajectory preprocessing → context construction → API-based LLM/RAG tool calls → evaluation.
  • Ran feasibility experiments, analysed failure cases and documented directions the team later optimised.

Final system on BW team outcome

Top-1 accuracy ↑ higher is better

Before60.05%
Final71.42%

Mean location error ↓ lower is better

Before3.44 km
Final2.20 km
Error axis scaled to 4 km.
  1. 01Trajectory preprocessing
  2. 02Context construction
  3. Routine case History-based fast path
    Ambiguous case Tool-assisted LLM reasoning recent contexthistorical behaviourstay / move likelihoodcandidate profiles
  4. 03Next-location prediction
Two-path design: a cheap lookup where behaviour is regular, an evidence-gathering agent where it is not.

Experience

Research and industry, across Singapore and Japan

  1. Institute of Data Science (IDS), National University of Singapore

    Research Assistant · Mentor: Prof. Ng See Kiong

    Sep 2026 – PresentSingapore

    • Contribute to research on spatiotemporal data mining and smart-city (Simulation & AI) applications.
  2. LocationMind Inc.

    Development Intern

    Jul 2025 – Mar 2026Tokyo, Japan

    • Built ETL inputs for multidimensional geospatial datasets, covering ingestion, cleaning, transformation and analysis reporting.
    • Joined urban, socioeconomic and infrastructure layers to produce commercial POI location features; compared mobility-prediction and spatial-analytics methods for product and architecture reviews.
    • Benchmarked large multimodal models on spatial reasoning with satellite imagery and geographic context.

    Geospatial ETLPOI featuresLMM benchmarking

  3. Information Technology Center (ITC), The University of Tokyo

    Collaborative Research Assistant · Mentor: Prof. KOBAYASHI-HILL Hiroki

    Mar 2025 – Aug 2025Chiba, Japan

    • Drove early-stage research on an LLM-agent system for human mobility prediction — from scoping the research question and designing the initial workflow to prototype implementation and feasibility experiments.
    • Reviewed and presented recent literature in lab seminars, synthesised research directions, and contributed to cross-disciplinary collaborations spanning human mobility, LLM agents and medical AI.
    • Independently collected and structured 1.8 million Japanese horse-racing records covering ~270,000 races, and constructed a pedigree network with over 100,000 nodes.

    LLM agentsHuman mobilityNetwork analysis

Selected projects & publications

Cities, agents and networks

Case competition · LIFT 2026 · Team project

UrbanPulse: who does a plan cool?

Block-level heat-resilience and inclusion index for Singapore's 10,749 HDB blocks — six pillars from 89 Landsat scenes, Sentinel-2, a shade-mapped OSM walking network and 16 data.gov.sg datasets, weighted by where seniors and rental households live — plus a green-finance business model.

Remote sensingNetwork analysisGurobi MILPMapLibre

Oct 2026 · with Liu Qiran

Publication · ACM HIDA 2026

MACA-PA: Modular Agent-Based Clinical Assistant for Perioperative Anesthesia

A locally deployable multi-agent, RAG clinical assistant: a root agent routes tasks to specialised agents for clinical summaries, CRF generation and medical-history classification.

My part Publication-ready workflow diagrams (task routing, local-RAG prompt assembly, agent coordination); edited and rewrote manuscript sections for technical and clinical consistency.

Jiang, Y., Ma, C., Xu, Z., Mu, D., Ni, L., Kobayashi, H. Proceedings of HIDA 2026, ACM.

ACM Digital Library ↗

Nov 2025 – Jan 2026

Research · SUSTech

Scientific Collaboration Network Analysis

Used embedding-based representations to explore potential gender differences in research-field patterns among researchers with comparable research performance.

EmbeddingsNetworkXPython

Sep 2024 – Jan 2025

Research · The University of Hong Kong

LLM-Based Entity Extraction & Collaboration Networks

Used an LLM to extract entities and relationships from COVID-19-era news reports and constructed a pandemic-response collaboration network.

LLM extractionNetwork analysis

Jul 2024 – Sep 2024

Education & skills

Statistics foundation, data-science focus

  1. Aug 2026 – Jul 2027 (expected)

    National University of Singapore

    MSc in Data Science for Sustainability

  2. 2022 – 2026

    Southern University of Science and Technology

    BSc in Statistics · Shenzhen, China

    Outstanding Thesis — Machine Learning-Based Community Detection in Scientific Collaboration Networks

Data analytics
PythonpandasNumPyNetworkXSQLETLFeature constructionModel evaluation
Geospatial & mobility
QGISSpatial joinsGPS / POI analysisSpatiotemporal featuresMobility data
AI agents
LLM workflowsTool callingRAGPrompt designBenchmarking
Communication
Workflow diagramsTechnical writingResearch synthesisAnalytical reporting
Languages
Chinese · nativeJapanese · JLPT N1English · IELTS 6.5

Contact

Looking for someone who connects data, cities and applied AI?

Open to data analytics, applied AI and smart-city / geospatial roles in Singapore.