Goh Kun Ming

Data · Research platform

Policy Intelligence

Mar 2025 – Present

Historical data, connected for careful exploration.

A reproducible platform for historical security and public-policy research, combining data pipelines, machine learning, geospatial analysis, graph views and guarded retrieval. A FastAPI backend and React dashboard support aggregate exploration. The project reports associations rather than causal policy effects, and does not claim model quality without the licensed data and pinned results.

• FastAPI
• React
• Data Engineering
• RAG

Inside this build

• Built a 1996–2021 country-year panel connecting incident outcomes with governance indicators.
• Implemented adaptive casualty-severity labels with clustering and deterministic fallback rules.
• Configured Logistic Regression, Random Forest and Extra Trees within shared preprocessing and experiment tracking.
• Authored 12 research notebooks covering data quality, forecasting, models, graphs, retrieval and monitoring.
• Added bounded API responses and safeguards that restrict use to historical aggregate research.

Research · Public project

Cross-Site RF Drone Detection and Classification

May 2026 – Aug 2026

Research on cross-site radio-frequency classification, with documented evaluation limits and no deployment claim.

A research study examining whether radio-frequency drone classifiers generalise across collection sites. It compares signal representations and separates drone detection from known-identity classification. The reported evaluation did not meet every cross-site gate, so location-independent performance, publication readiness and deployment are not claimed.

Inside this build

• Audited 1,253 active captures across four sites and quarantined invalid sessions.
• Converted captures into consistent signal windows and PSD/STFT representations.
• Preregistered held-site evaluation, thresholds and acceptance criteria without tuning on the held-out site.
• Compared PSD, STFT and combined representations with clustered uncertainty estimates.
• Evaluated known-identity and unknown-identity cases separately.
• Documented failed acceptance gates and limits on deployment and generalisation.

Research · Public project

Multi-Modal Low-Altitude Drone Detection and Digital Twin

Apr 2026 – Aug 2026

Event-camera detection research and sensor-interface work; the public record distinguishes offline experiments from live system readiness.

An event-camera detection study alongside event, thermal and RF sensor-interface integration and a separate RGB detector. The reported model results apply to the offline event-camera branch; the other sensors were not learned fusion inputs. Live-camera readiness and an operational fused system are not established.

Inside this build

• Integrated three sensor interfaces and documented model inputs, outputs and handover limitations.
• Converted 231 FRED sequences into 630,340 causal event-stack frames with sequence-separated splits.
• Audited annotations, images, projected boxes and tiles before evaluation.
• Compared YOLO and RF-DETR training runs and stitched-image configurations.
• Selected an operating point that traded some detection accuracy for fewer false positives.
• Documented offline GPU timing, rain-filter limitations and missing live-system validation.

Research · Public project

Satellite-to-Simulation Data Pipeline

Apr 2026 – Aug 2026

A research pipeline for harmonising imagery annotations, tracking provenance and checking segmentation data quality.

A governed satellite-imagery pipeline that converts heterogeneous annotations into segmentation, classification and context outputs. The work combines consistent label mapping, geometry-derived features and traceable quality checks to prepare data for simulation-related research.

Inside this build

• Harmonised raster, COCO, XML, TXT and GeoJSON annotations without silently accepting ambiguous labels.
• Prepared 13,487 segmentation samples and 56,096 classification rows across 14 dataset roots.
• Standardised one background and 19 foreground classes.
• Derived geometry components, topology edges and context features.
• Produced provenance logs, structured inventories, stage reports and 117 visual QA previews.
• Checked image-mask alignment and duplicate-image leakage, with unreadable data quarantined.

Research · Public project

Tri-Modal Runway Inspection and Abnormality Detection

Apr 2026 – Aug 2026

A simulated inspection study comparing sensor and fusion strategies, with explicit limitations on operational use.

A simulated runway-inspection study comparing RGB, thermal and LiDAR detection and fusion strategies. UE5/AirSim samples support reproducible experiments and provenance review. Reported pre-audit diagnostics exposed detection failure, so the work is presented as a research study rather than a deployable inspection system.

Inside this build

• Captured 1,442 simulated samples across 14 conditions with crack and pothole annotations.
• Kept model-based thermal outputs distinguishable from rule-derived detections.
• Compared single-sensor, pairwise and tri-sensor fusion strategies.
• Separated training and held-out sequences, with tuning restricted to training data.
• Estimated uncertainty with sequence-clustered bootstrap analysis.
• Tracked raw artefacts and routing decisions, while recording the limits revealed by evaluation.

AI · Public project

YOLO Leaf Object Detection Pipeline

Jul 2026

Dataset preparation and validation for leaf detection; prepared data does not establish trained-detector performance.

A reproducible leaf-detection workflow combining PlantVillage, PlantDoc and generated hard negatives. It prepares and validates data, configures YOLO training and provides export and browser-inference hooks. No committed trained weights or held-out results establish detector accuracy, latency or production readiness.

Inside this build

• Prepared 57,164 images and 63,225 boxes in deterministic training, validation and test splits.
• Converted Pascal VOC annotations to normalised YOLO coordinates.
• Generated collision-resistant filenames across source datasets.
• Validated image-label pairs, class IDs, coordinates and split leakage before training.
• Configured three YOLO candidates, a bounded smoke-training path and ONNX export hooks.
• Kept dataset-tool tests separate from external-data preparation and GPU training.

Security · Public project

Sentinel Endpoint Framework (SEF)

Jun 2025 – Jun 2026

A local endpoint watchdog with physical-key checks, PIN controls, audit records and user-controlled recovery.

A Rust-based local endpoint watchdog and control dashboard using physical-key presence, PIN-protected controls and audit records. It supports reversible maintenance and user-controlled recovery, with a Python fallback during migration. Advanced protection modules remain preview designs beyond the implemented user-mode watchdog.

Inside this build

• Implemented key-presence polling, grace periods and fail-safe locking requests.
• Added salted PIN verification, short authorisation sessions and lock-first recovery requiring the physical key.
• Built a localhost dashboard for trusted-key enrolment, protection state and maintenance controls.
• Recorded key transitions, configuration changes, authorisation and recovery events.
• Packaged reversible Windows startup controls and bounded long-running policy state.
• Separated seven preview control families from implemented features through an explicit threat model.

AI · Full-stack prototype

VeggieAI

Jan 2026 – Feb 2026

Computer vision meets a useful web application.

VeggieAI connects vegetable image classification with price lookup, authenticated history and conversational guidance. A Flask web application uses a separate registry-backed model service. Two reproducibly loadable model packages are documented, while held-out accuracy and latency claims remain unsupported by committed training and evaluation artefacts.

• Python
• Flask
• Computer Vision
• Docker

Inside this build

• Separated the web application and inference service into a two-image container stack.
• Defined distinct four-label and five-label model contracts with model-specific preprocessing.
• Validated manifests, labels, package paths and checksums before serving.
• Rejected incompatible image shapes, channels and normalisation profiles before inference.
• Exercised health, metadata and prediction interfaces through service smoke tests.
• Kept OpenAI, Roboflow, email and remote-model integrations optional and explicitly configured.

Community · Public project

RAG-Powered Food Waste Reduction System

Oct 2025 – Feb 2026

A campus food-waste chatbot concept developed through design thinking, vendor insights and feasibility analysis.

FoodBot is a campus food-waste chatbot concept developed through design thinking, vendor insights and feasibility analysis. It explores preparation guidance for vendors and better disposal decisions for students and staff, using time, location, waste patterns and academic-calendar context. It is a proposal, not a deployed service or a measured waste-reduction outcome.

Inside this build

• Analysed campus food-waste patterns and operational gaps in existing approaches.
• Considered vendors, cleaners and consumers when framing the system-level problem.
• Applied the S.P.I.C.E. framework to investigate behavioural drivers.
• Developed personas and a low-effort chatbot interaction concept.
• Outlined preparation guidance, disposal prompts and behavioural nudges.
• Connected the proposal to responsible consumption and campus sustainability goals.

Data · Coursework dashboard

Student Risk Analytics

Jan 2026 – Feb 2026

Making cohort patterns easier to understand.

An interactive Dash application exploring student performance, attendance, surveys and cohort patterns. It combines source data into latest-semester snapshots and transparent review views. Configurable GPA and attendance thresholds provide non-diagnostic decision-support heuristics; they do not establish causal risk factors or effective interventions.

• Python
• Dash
• Statistics
• Visualisation

Inside this build

• Merged student profiles, results, surveys and course codes, and unified period labels for filtering.
• Created latest-semester snapshots and three configurable review tiers.
• Built cohort, trend, risk-flow, support-factor and priority views.
• Exposed filters and thresholds so reviewers can inspect how groupings change.
• Checked workbook contracts, identifier overlap, missing data and duplicates.
• Documented reproducible application configuration and validated analytical helpers.

AI · Public project

Trip Safety Prediction System

Jan 2026 – Feb 2026

An offline trip-classification application that distinguishes model-backed batch predictions from single-record heuristic outputs.

An offline Windows application that converts vehicle sensor records into trip-level safety classes using a bundled decision-tree pipeline. Batch predictions use a defined feature contract, while single-record outputs are clearly labelled as heuristic fallbacks. Reproducible inference is demonstrated; held-out predictive quality is not established by the committed material.

Inside this build

• Validated missing-value markers, GPS accuracy, bearings and required numeric fields.
• Derived 19 trip-level features covering duration, movement, speed and sensor quality.
• Produced model-backed SAFE or DANGEROUS batch outputs with inference-source labels.
• Separated full-trip aggregation from single-record heuristic mode.
• Packaged the model and interface assets into a Windows executable.
• Checked cleaning, aggregation, persistence, desktop flows and packaged inference through automated tests.

Data · Public project

Wage Modelling & Regression Analysis

Oct 2025 – Feb 2026

A regression coursework study combining analytical derivation, gradient descent and reproducibility checks.

A transparent regression study using wage data and a small derived overtime example. It connects analytical derivation, from-scratch gradient descent, feature scaling and statistical cross-checks. Reported improvements concern computational convergence, not predictive generalisation; the two-feature example contains only ten derived rows.

Inside this build

• Implemented through-origin, intercept-and-slope and two-feature regression models in Python.
• The LinkedIn project description reports comparing ten shared learning rates through 20 deterministic fits.
• The LinkedIn project description reports convergence reductions from 173 to two iterations and from 401 to 59 within stated error tolerances.
• Cross-checked fitted objectives against closed-form least-squares references and Minitab.
• Added reproducibility and guardrail tests with documented derivations and reporting sources.

Data · Public project

Driver Safety Analytics Platform

Oct 2025 – Dec 2025

A driver-safety data pipeline combining validation, feature engineering, dimensional modelling and analytical reporting.

A driver-safety data-engineering pipeline moving driver, sensor and safety-label records from MySQL staging through Python and PySpark processing into a star-schema warehouse. It prepares data and Power BI artefacts for exploratory analysis, with safe command-line orchestration and offline checks rather than claims of production scale or safety impact.

Inside this build

• Modelled driver and trip-sensor dimensions linked to booking-level safety labels.
• Applied schema standardisation, deduplication, missing-value handling and rule-based validation.
• Derived acceleration magnitude, jerk, trip aggregates and behaviour indicators.
• Implemented IQR and PCA-reconstruction diagnostics for unusual sensor records.
• Added environment validation, dry runs and opt-in live database access.
• Preserved notebooks, SQL, Power BI artefacts and architecture, privacy and operations documentation.

Data · Coursework dashboard

HDB Resale Explorer

Oct 2025 – Dec 2025

Exploring the data behind a Singapore home.

A Tableau-ready market-analysis pipeline exploring Singapore HDB resale affordability, value and location trade-offs. Reproducible cleaning and geospatial enrichment feed validated datasets and a packaged dashboard. Market measures retain their stated time periods, keeping year-over-year changes distinct from longer-term growth.

• Python
• Data Cleaning
• Dashboards
• Singapore

Inside this build

• Standardised town, street, flat-type, storey, model and lease labels.
• Corrected documented floor-area anomalies and removed duplicate records.
• Validated canonical and enriched datasets containing 202,461 and 216,695 rows.
• Derived price-per-area, lease, transaction-period and location-context fields.
• Built a Tableau workbook for comparisons across towns, flat types and household needs.
• Preserved notebooks, methodology, data-source notes, reports and presentation artefacts.

Inside the project

3 images

AI · Public project

Multimodal Property Price Prediction System

Oct 2025 – Dec 2025

A property-valuation prototype combining structured listings with optional text and image signals, with bounded component evaluation.

EstateScope AI is a Flask-based US property-valuation prototype combining structured listings with optional description and image signals. It supports prediction history, analytical interfaces, housing chat and model governance. The text component has a bounded held-out evaluation, while comparable evidence for multimodal uplift or production accuracy is not established.

Inside this build

• Prepared leakage-separated text training, validation and test sets.
• Evaluated the text component on 1,517 held-out examples with a pinned model artefact.
• Exposed per-modality estimates and averaged only inputs available at inference.
• Verified committed model checksums, loading and compatibility.
• Built bounded trend and distribution interfaces with authenticated history.
• Documented model promotion, rollback, container health and missing multimodal evaluation evidence.

Creative · Public project

Data Storytelling on AR Gaming Pilot with Current Affairs Insights

Jul 2025 – Aug 2025

Power BI dashboards and a team presentation using demographic and gaming insights to discuss pilot locations.

A Power BI data story and team presentation proposing locations for an augmented-reality game pilot. Demographic, accessibility and gaming insights inform the recommendation, with selected Sustainable Development Goals included as coursework context.

Inside this build

• Developed a persuasive data story from an earlier analysis.
• Refined Power BI dashboards, visual cues and narrative flow.
• Compared demographic, accessibility and gaming-behaviour evidence for Catch & Go.
• Delivered a 15-minute team presentation and answered stakeholder questions.
• Recommended Singapore pilot locations and connected the proposal with selected Sustainable Development Goals.

Creative · Public project

Diploma in Applied AI & Analytics Promotion Video

Jul 2025 – Aug 2025

A diploma promotion video exploring audience-focused storytelling, platform selection and responsible digital content creation.

A short promotional video introducing the Diploma in Applied AI & Analytics to audiences aged 16–21. The project combines audience-focused storytelling, platform selection and responsible digital-content creation.

Inside this build

• Produced a video showcasing the diploma for a youth audience.
• Applied PAC and HoMA frameworks to develop the story and opening hooks.
• Evaluated social-media platforms for the intended audience.
• Considered copyright, attribution and responsible AI-tool use.
• Explained the creative choices, outreach approach and digital ethics in a Q&A session.

AI · Notebook experiment

Handwritten Character Lab

Jul 2025 – Aug 2025

Comparing the ways neural networks create.

A generative-modelling lab for 16 selected EMNIST Letters classes, preserving coursework notebooks alongside reusable Python tools. It explores eight GAN families and variational or conditional extensions. Evaluation tools are provided, but committed checkpoints and multi-seed results do not establish a winning model.

• Deep Learning
• GANs
• EMNIST
• Evaluation

Inside this build

• Configured consistent grayscale inputs, latent vectors and reproducible experiment settings.
• Implemented eight GAN families alongside variational and conditional extensions.
• Added FID, KID, diversity, mode-collapse and embedding-analysis tools.
• The LinkedIn project description reports validating sample counts, finite values, dimensions and covariance structure before computing metrics.
• Documented limitations of natural-image feature extractors for small handwritten letters.
• Preserved the original notebook and output-free research sections with preprocessing and integrity tests.

Software · Public project

Newspaper Restoration System using Prefix Tries

Jul 2025 – Aug 2025

A text-restoration tool exploring wildcard recovery, fuzzy suggestions, vocabulary maintenance and trie visualisation.

An explainable Python command-line toolkit for restoring incomplete historical newspaper text. Each wildcard represents one missing character, with prefix-trie lookup, ranked suggestions and fuzzy search supporting recovery. Visual inspection and automated checks explain behaviour without claiming measured restoration accuracy on a benchmark corpus.

Inside this build

• Implemented exact wildcard, subtree and edit-distance recovery strategies.
• Ranked exact and fuzzy suggestions deterministically.
• Added prefix and suffix batch editing with persistent vocabulary storage.
• Visualised trie topology, similarity relationships and co-occurrence structure.
• Separated storage, file handling, restoration, graphs and console interaction into testable components.
• Provided CLI workflows for editing, search, batch processing, visualisation and recovery.

AI · Notebook experiment

Pendulum Learning Lab

Jul 2025 – Aug 2025

Learning control through trial, reward and evaluation.

A reinforcement-learning lab comparing Standard, Double, Dueling and Rainbow DQN on Pendulum-v1. Continuous torque is mapped to discrete actions, with reusable replay, training, evaluation and tuning components. Local smoke execution is kept distinct from full studies; matched multi-seed reward superiority and control stability are not claimed.

• Reinforcement Learning
• DQN
• Python
• Optimisation

Inside this build

• Mapped continuous torque to discrete action bins for value-based learning.
• Configured seeded experiments with replay buffers, warm-up and target-network updates.
• Added prioritised replay and distributional outputs for Rainbow DQN.
• Provided reward summaries, learning-curve analysis, greedy rollouts and comparison views.
• Separated bounded smoke runs from the full 500-episode study configuration.
• Validated experiment tooling and notebook integrity while documenting missing matched multi-seed runs.

Creative · Public project

AI-Powered Haiku Generator

Apr 2025 – May 2025

A command-line haiku tool combining word transformations, Markov-chain generation, syllable checks and optional audio narration.

A command-line haiku tool combining word transformations, syllable-aware Markov generation and optional narration. The project explores how explicit language rules and bounded generation can produce a reproducible 5–7–5 poem.

Inside this build

• Built synonym, shortest-synonym, antonym and batch-permutation modes with punctuation and capitalisation handling.
• Implemented syllable-aware Markov generation with bounded retries and fallbacks for the 5–7–5 structure.
• Added optional WAV narration assembled from word-level audio clips.
• Separated tokenised lines, keyword indexing, syllable counting, thesaurus lookup, haiku generation and file handling into six abstractions.
• Validated command-line input and file availability, and added 16 tests with 54% coverage.
• Recorded a deterministic 5–7–5 example from a 12-line corpus; this demonstrates the generation path rather than general poetic quality.

AI · Public project

CNN Image Classification on Multi-Resolution Vegetable Dataset

Apr 2025 – May 2025

A coursework comparison of convolutional architectures, image resolutions and augmentation for vegetable classification.

A vegetable-image classification study comparing convolutional network designs across two image resolutions and augmentation settings. It combines dataset repair, repeatable preprocessing and evaluation of historical experiment results.

Inside this build

• Compared 11 vegetable classes across 23×23 and 101×101 grayscale inputs, with and without augmentation.
• Corrected validation/test directory names and misplaced carrot images using helpers with a dry-run mode.
• Explored Sequential, Functional, residual, Inception-style and depthwise-separable CNN families.
• Documented a historical Sequential search with 0.908 validation accuracy and a separate run with about 94.2% test accuracy; the results do not establish a universal architecture winner.
• Added 13 tests across Python 3.11 and 3.12, with 48% broad coverage.
• Recorded dataset split counts and missing reproducibility artifacts, including model weights and multi-seed comparisons.

Data · Public project

Location Recommendation for AR Mobile Game Pilot

Apr 2025 – May 2025

A Power BI analysis of demographics, accessibility and gaming behaviour to inform a proposed pilot location.

A Power BI location-recommendation project for Funtendo’s proposed Catch & Go augmented-reality game pilot. The team used demographic, mobility and gaming information to explain candidate locations in Singapore.

Inside this build

• Combined demographic, income, dwelling, mobility and gaming datasets in a Power BI analysis.
• Created visualisations to compare accessibility, player behaviour and engagement considerations.
• Proposed Singapore pilot locations using the patterns identified in the report.
• Summarised the recommendation in a concise key message and a Sustainable Development Goals justification.
• Presented the analysis through a three-page dashboard and supporting narrative.

AI · Public project

RNN-based Movie Review Sentiment Analysis

Apr 2025 – May 2025

A movie-review study comparing recurrent models for sentiment classification and rating prediction with separated evaluation data.

A bilingual movie-review study comparing recurrent networks for sentiment classification and rating regression. The workflow emphasises consistent English and Malay preprocessing, leakage-aware splits and task-appropriate evaluation.

Inside this build

• Created 80/10/10 training, validation and test splits before fitting the modelling workflow.
• Cleaned English and Malay reviews through lowercasing, punctuation handling, stop-word removal, lemmatisation, tokenisation and padding.
• Applied seeded augmentation only to training data, expanding classification examples from 222 to 888 and regression examples from 501 to 2,004.
• Compared SimpleRNN, LSTM and GRU models for binary sentiment and numerical rating prediction.
• Evaluated classification with accuracy, precision, recall and F1, and regression with error and fit metrics across rating bands without claiming an architecture winner.
• Added 22 tests with 55% coverage and documented missing-score, duplicate, short-review and repeated-text cleanup.

Research · Public project

Research Report on Quantum Computing as Green Technology for AI Model Training

Apr 2025 – May 2025

A literature-based discussion of quantum computing, AI training and sustainability; it is not a measured energy-efficiency benchmark.

A research report exploring quantum computing as a possible green technology for AI training. It examines energy use, cost, technical limits and adoption options through published literature rather than a workload-matched experimental benchmark.

Inside this build

• Reviewed literature on quantum and classical computing for AI model training.
• Compared reported energy, infrastructure and cost considerations alongside technical limitations.
• Organised opportunities and risks through a SWOT analysis and environmental-impact discussion.
• Connected the report to SDGs 9, 12 and 13.
• Recommended investigating hybrid and cloud-based adoption while treating sustainability benefits as questions requiring comparable workload evidence.

Community · Public project

NUS HumanITy Challenge

Mar 2025 – Apr 2025

A sustainability innovation challenge involving technology ideas, risk considerations and collaborative solution design.

A finalist team experience in the NUS HumanITy Challenge, a pre-university sustainability competition. The project focused on developing an IT- and AI-based response to a real-world sustainability problem.

Inside this build

• Developed sustainability solution ideas using information technology and AI.
• Considered implementation risks while refining the proposed solution.
• Collaborated across disciplines to present the team’s finalist proposal.

Data · Public project

Content Performance Analysis & Strategy Refinement

Jan 2025 – Feb 2025

An exploratory analysis of campaign formats, posting times and engagement patterns, with reproducible figures.

An observational analysis of social-content performance, using historical post data to refine a proposed digital-marketing strategy. The project separates measured engagement patterns from recommendations that had not been tested in a subsequent campaign.

Inside this build

• Analysed a corpus of 100 posts, 192 comments, 689 reactions and 8,208 views, with a focused review of seven campaign pieces over 28 days.
• Compared content formats, posting days and times against views, reactions and comments.
• Used heatmaps, box plots, line charts and grouped comparisons to explain engagement patterns.
• Proposed changes to posting schedules, calls to action, hashtags and content based on the observed data.
• Added 14 regression-contract tests and checked nine reproduced figures for byte-identical output.
• Kept findings observational; the analysis does not measure causal effects or post-campaign uplift.

AI · Notebook experiment

Customer Segmentation

Jan 2025 – Feb 2025

Exploring customer groups through clustering.

A customer-segmentation workflow comparing clustering approaches for age, income and spending data. It includes input validation, reusable model artifacts and interpretable segment summaries for exploratory marketing analysis.

• Python
• Clustering
• PCA
• Evaluation

Inside this build

• Validated finite, unique customer identifiers and documented categorical values, with a 1.5×IQR income-outlier rule.
• Standardised age, income and spending features while reserving gender for post-clustering interpretation.
• Compared K-Means, agglomerative clustering and DBSCAN, with and without PCA.
• Built a configurable six-segment K-Means workflow that saves its scaler, model, assignments and metrics.
• Evaluated silhouette, Davies–Bouldin and Calinski–Harabasz scores alongside inertia, and added 15 tests with 95% coverage.
• Documented the absence of external validation, stability studies and measured campaign lift; the segments remain exploratory.

Software · Public project

FitnessQuest Gamified Fitness Application

Oct 2024 – Feb 2025

A gamified fitness application with accounts, challenges, quests, inventory and ownership-aware application controls.

A responsive gamified fitness application built with Express, MySQL and a vanilla JavaScript interface. It connects user accounts, challenges and game progression with explicit ownership checks and repeatable application tests.

Inside this build

• Built registration, login, profiles, leaderboards, challenges, reviews, pets, inventory, battles, quests and achievements.
• Implemented an Express 5 application for Node.js 22+, backed by MySQL 8.4, JWT authentication and Zod validation.
• Enforced ownership using the authenticated JWT principal instead of trusting client-supplied user identifiers.
• Added password hashing, security headers, request limits, parameterised queries and safe error responses.
• Rendered user content with textContent and separated database migrations and test resets from the application runtime.
• Documented six API, three security and six browser checks; these isolated checks do not establish production reliability.

Data · Public project

Singapore Graduate Employment, Salary & Labour-Market Analysis

Jan 2025 – Feb 2025

A graduate-employment and salary study using data-quality checks, exploratory modelling and transparent group comparisons.

A Singapore graduate-employment and labour-market analysis combining data cleaning, descriptive views, salary modelling and clustering. The project examines employment and income patterns across education and population groups.

Inside this build

• Prepared a workflow for seven source datasets, covering type conversion, missing values, duplicates and IQR or z-score outlier checks.
• Derived salary-spread and employment-rate-gap measures and mapped fields of study.
• Used random forests for salary imputation or prediction alongside clustering and descriptive analysis.
• Compared patterns across universities, degrees, years, industries, age, gender, education and population context.
• Checked seven notebooks without stored errors and added 17 tests with 80% coverage across Python 3.10–3.12.
• Documented missing raw source tables, fitted models, held-out metrics and uncertainty analysis; the results do not establish predictive performance.

Data · Public project

Statistical Analysis of AI Industry Trends

Nov 2024 – Feb 2025

A statistical coursework investigation of AI-industry trends, presented through charts and a data-storytelling poster.

A statistical investigation of AI-industry trends, presented as a data-storytelling poster. The work examines relationships involving training compute, model performance, experience, salaries and AI adoption.

Inside this build

• Explored real-world datasets describing the AI industry.
• Investigated relationships between training compute and model performance, and between experience and salary.
• Compared salary patterns associated with AI adoption using descriptive statistics and group analysis.
• Applied ANOVA and Tukey HSD alongside correlation and regression views.
• Communicated the findings through an A3 poster with box plots, histograms and supporting statistical explanations.

AI · Notebook experiment

Utilities Forecasting

Jan 2025 – Feb 2025

Studying seasonal patterns in electricity, gas and water use.

A monthly utilities forecasting study covering gas, electricity and water consumption. The workflow compares statistical time-series models with chronological evaluation and validates the structure of the input series.

• Python
• Time Series
• ARIMA
• Evaluation

Inside this build

• Prepared monthly series with log transformation and chronological holdouts, documenting how optional outlier removal can affect temporal spacing.
• Compared ARIMA, tuned ARIMA and seasonal SARIMAX models.
• Recorded selected gas ARIMA(18,2,9), electricity SARIMAX(3,1,4)(5,1,6,12) and water ARIMA(9,2,17) configurations.
• Structured a 60-month holdout and forecast horizon with MAE, RMSE, R², explained variance and MAPE evaluation.
• Validated finite values, contiguous monthly dates and forecast horizons, rejecting duplicate or missing months; added 43 tests.
• Documented missing raw series, fitted models, pinned backtests, baseline comparisons and forecast intervals; no operational or sustainability impact is established.

Community · Public project

NUS-Rightship Maritime Hackathon

Oct 2024 – Jan 2025

A maritime hackathon exploring AI and data-driven ideas for port operations, emissions and vessel safety.

A team project for the 24-hour NUS–RightShip Maritime Hackathon 2025. The challenge involved developing data- and AI-informed ideas for problems in global maritime operations.

Inside this build

• Worked on a maritime challenge under a 24-hour hackathon format.
• Explored AI and data-driven approaches to port efficiency, emissions and vessel safety.
• Applied machine learning, analytics and collaborative problem-solving to the proposal.

Research · Public project

AI Ethics & Governance Survey Analysis

Oct 2024 – Nov 2024

A survey-based exploration of perceptions of AI legislation, transparency, data protection and governance.

A survey-analysis project examining public and expert perspectives on AI ethics, legislation and governance. A Power BI dashboard communicates respondents’ concerns and priorities in relation to SDG 16.

Inside this build

• Analysed survey responses on AI legislation, ethics and governance.
• Examined transparency, international law, data protection, misinformation and job displacement.
• Used descriptive statistics to compare and rank respondents’ concerns.
• Built Power BI views of responses and governance priorities.
• Proposed greater transparency, international standards and privacy protections based on the survey findings.

Creative · Public project

Digital Marketing Strategy for Bubble Tea Brand

Oct 2024 – Nov 2024

A marketing coursework proposal using environmental analysis, marketing-mix frameworks and customer-journey planning.

A digital-marketing strategy proposal for a bubble-tea brand. The work connects customer research, positioning and a two-week content plan around the value proposition “Less Stress, More Boba.”

Inside this build

• Analysed the business context using STEEP, SWOT and the marketing mix.
• Developed customer profiles and mapped the buyer decision process.
• Defined an online value proposition around custom drinks and flavoured pearls.
• Planned content pillars, platforms and audience targeting in a two-week editorial calendar.
• Outlined paid-promotion ideas and budgets as a proposal, without claiming measured campaign results.

Data · Coursework analysis

Job Posting Explorer

Oct 2024 – Nov 2024

Exploring the patterns behind job-posting engagement.

An exploratory LinkedIn job-listing analysis focused on salary, job attributes and application activity. The workflow cleans and joins source tables, derives comparison measures and documents how the notebook can be reconstructed.

• Python
• Pandas
• Matplotlib
• Data Cleaning

Inside this build

• Prepared five source tables covering companies, salaries, activity, work arrangements, pay periods, seniority and timestamps.
• Derived annual salary, salary range, posting duration and a zero-safe application-to-view ratio.
• Examined salary, work arrangements, seniority, posting duration, views, applications, annual pay and conversion patterns.
• Applied IQR salary filtering and optional z-score checks for outlier exploration.
• Organised a 234-cell notebook into six sections with read-only reconstruction verification and automated tests.
• Documented the unavailable raw source tables and kept conclusions within descriptive job-listing analysis.

AI · Notebook experiment

Machine Fault Prediction

Oct 2024 – Nov 2024

Comparing models for machine-status classification.

An ensemble-classification study for a strongly imbalanced machine-fault dataset. The workflow separates training, threshold selection and final evaluation, reporting fault-detection trade-offs rather than accuracy alone.

• Python
• Machine Learning
• Ensemble Learning
• Evaluation

Inside this build

• Prepared 20,000 records with 678 faults and 2,579 missing cells using training-fitted preprocessing.
• Combined decision tree, random forest, gradient boosting, AdaBoost, K-nearest-neighbour and logistic-regression estimators.
• Separated 12,000 training, 4,000 threshold-validation and 4,000 test records with a fixed random seed.
• Selected a 0.455 decision threshold by validation F1 across 181 candidates.
• Reported held-out precision 0.5932, recall 0.7721 and F1 0.6709, detecting 105 of 136 test faults with 72 false positives.
• Added 17 tests and documented the single-dataset, non-temporal evaluation limit; external-factory validation and calibration remain unestablished.

AI · Notebook experiment

Housing Price Lab

Oct 2024 – Nov 2024

Examining housing-price models against a simple baseline.

An ensemble-regression study using a 545-row classroom housing dataset. The project engineers property features, compares model weights with cross-validation and reserves a separate test set for evaluation.

• Python
• Regression
• Ensemble Learning
• Evaluation

Inside this build

• Prepared a dataset spanning five cities and three renovation categories, with no missing values or duplicate identifiers.
• Engineered total-room, area-per-bedroom, toilet-to-bedroom and storey-area predictors while excluding identifiers and price from inputs.
• Used fold-fitted imputation, scaling and one-hot encoding to prevent preprocessing leakage.
• Compared four ensemble-weight candidates with five-fold validation on 436 training rows, selecting a 2:1:1 Ridge, random-forest and gradient-boosting ensemble.
• Evaluated 109 held-out rows: R² 0.5353, MAE 113,249.95 and RMSE 153,267.51, with lower error than the median baseline.
• Added 19 tests and documented the lack of external validation, prediction intervals and subgroup analysis; this remains a classroom study.

Data · Public project

Water & Energy Consumption Analysis of Shopping Mall

Oct 2024 – Nov 2024

A Power BI dashboard exploring shopping-mall water and electricity use, trends and category-level comparisons.

A Power BI analysis of five years of water and electricity consumption at Prestige Mall. The dashboard supports discussion of resource-use patterns and possible efficiency improvements.

Inside this build

• Analysed five years of shopping-mall water and electricity data.
• Built interactive views of consumption trends, categories and year-on-year changes.
• Used treemaps, line charts, bar charts and waterfall charts to explain the findings.
• Connected the analysis to SDG 6 and SDG 7.
• Proposed efficiency improvements through data storytelling without claiming measured resource savings.

Security · Public project

GovTech AI Capture-The-Flag

Oct 2024

Participation in an AI security competition involving team coordination and challenge-based problem solving.

A team participation in GovTech’s Singapore AI Capture-The-Flag competition. The experience combined technical problem-solving with coordination across a range of AI and cybersecurity challenges.

Inside this build

• Participated in GovTech CSG’s Singapore AI Capture-The-Flag competition.
• Tackled jeopardy-style challenges across seven domains.
• Led team coordination and collaborative problem-solving under competition conditions.

Research · Public project

Fairness of Meritocracy in Singapore

Apr 2024 – Aug 2024

A researched debate examining education access, examinations, scholarships and the strengths and limits of meritocracy.

A researched debate on fairness and meritocracy in Singapore’s education system. The work develops a structured case while examining the strengths and limits of examinations, scholarships and lifelong learning.

Inside this build

• Researched how meritocracy operates through Singapore’s education system.
• Examined examinations, scholarships and lifelong-learning opportunities using cited sources.
• Prepared a proposition case with supporting examples and structured arguments.
• Developed counterarguments and rebuttals addressing the limits of meritocratic outcomes.

Community · Public project

Resolution to Unequal Education Opportunities

Jun 2024 – Aug 2024

A proposal examining online learning as a response to unequal education access, including feasibility and digital-access limitations.

A proposal addressing unequal educational opportunities through a free online study-resource platform. The work explores socioeconomic barriers and evaluates the practical limits of a digital-access approach.

Inside this build

• Investigated socioeconomic barriers to educational opportunities.
• Proposed a free platform for study resources, notes and specimen papers.
• Applied SCAMPER and computational-thinking methods, including decomposition, pattern recognition, abstraction and algorithm design.
• Mapped the proposed workflow with a flowchart.
• Considered feasibility, scalability, digital access and infrastructure needs without claiming implemented educational outcomes.

Cloud · Public project

Website Deployment on AWS EC2 with WordPress

Apr 2024 – Aug 2024

A cloud coursework deployment of WordPress using an Ubuntu server, a LAMP stack and database configuration.

A cloud-hosting coursework exercise deploying WordPress on an Ubuntu AWS EC2 instance. The project covers server setup, networking, database configuration and web-application installation.

Inside this build

• Provisioned an Ubuntu 22.04 EC2 instance and installed a Linux, Apache, MySQL and PHP stack.
• Configured administrative SSH access through PuTTY and allocated an Elastic IP.
• Created the MySQL database and user privileges needed by WordPress.
• Configured Apache, PHP and application filesystem permissions.
• Tested the coursework site through its public IPv4 DNS at the time; no current live availability is claimed.

Research · Public project

Application of Artificial Intelligence in Driving

Apr 2024 – Jun 2024

A research proposal discussing AI for driving safety, alongside cost, fairness, privacy and security considerations.

A research presentation exploring possible AI applications for driving safety in support of SDG 3.6. The proposal weighs technical possibilities against cost, bias and security considerations.

Inside this build

• Researched AI approaches that could support road-safety and accident-prevention goals.
• Discussed autonomous systems, live monitoring and predictive analytics as potential approaches.
• Explained the roles of computer vision, machine learning, neural networks, deep learning and supporting data.
• Compared possible benefits with cost, bias and security challenges.
• Proposed responsible-deployment considerations, including algorithmic fairness; the work does not claim an implemented vehicle or measured safety improvement.

Community · Public project

NTU-GAIP Insurance Case Competition 2024

Jun 2024

An insurance case-competition proposal exploring AI support, fraud detection and pricing with responsible-governance considerations.

An AI-focused proposal for the NTU–GAIP Insurance Case Competition 2024. The case considers how customer support, fraud detection and premium-setting could be improved while accounting for responsible governance.

Inside this build

• Researched AI opportunities for an insurance case-competition proposal.
• Explored customer-support, fraud-detection and dynamic-premium applications.
• Considered implementation feasibility alongside ethical requirements.
• Addressed privacy, bias and responsible governance in the proposed approach.

Research · Awareness proposal

Youth Awareness Proposal

Oct 2022 – Nov 2022

A Boys’ Brigade presentation about youth awareness and community vigilance.

A Boys’ Brigade Global Awareness presentation proposing a youth workshop on radicalisation risks, digital literacy and community vigilance. The work records a researched prevention proposal rather than a delivered programme or measured impact.

• Research
• Digital Literacy
• Community
• Risk Awareness

Inside this build

• Researched online and offline radicalisation risks for a youth-awareness presentation.
• Outlined an educational workshop on recognising concerning content, seeking help and responding safely.
• Connected early-intervention themes with digital literacy and community vigilance.
• Considered sensitivity, delivery risks and possible mitigations.
• Planned roles, manpower, partnerships and resource needs for the proposed workshop.

Research · arXiv preprint · 2025

Quantum–Classical GANs

An experiment at the edge of two computing worlds.

My 2025 arXiv preprint explores noisy quantum circuits as latent priors in generative models, comparing classical and hybrid quantum-classical GANs.

• Quantum Computing
• Qiskit
• GANs
• Research

Inside this build

• Compared hybrid variants with 3, 5 and 7 qubits using Qiskit simulation.
• Evaluated generated binary-MNIST samples with FID and KID.
• Reported that the classical baseline performed best in these experiments.

Research · arXiv preprint · 2025

Residual Learning Study

Understanding why skip connections help deep networks.

A 2025 arXiv preprint co-authored with Xingyu Liu, studying residual connections through a ResNet-18 and conventional deep-CNN comparison on CIFAR-10.

• Computer Vision
• ResNet
• CIFAR-10
• Research

Inside this build

• Studied the role of residual connections in gradient flow.
• Compared training behaviour and classification results on CIFAR-10.
• Documented the implementation and evaluation in a public preprint.