EZAIRIEL

• slecare

  explainable clinical decision-support for lupus nephritis, built on real hospital data and read by clinicians rather than data scientists.

  0.80f1 · kidney damage

  grand award · accepted for publicationinternational journal of stem research · jun 2026

  • benchmarked catboost, random forest, svm, mlp, and stacked ensembles over demographic, laboratory, biopsy, and treatment data from hospital canselor tuanku muhriz, reaching f1 = 0.80 for kidney damage and f1 = 0.77 for delayed remission across baseline, 6-month, and 12-month horizons.
  • implemented a shap explainability layer over the selected model to isolate the dominant clinical risk factors behind each prediction and validate them against clinician domain expectations.
  • converted model probabilities into a clinician-facing dashboard of ranked, interpretable patient risk profiles.

  catboostshapstacked ensemblesclinical ml

• sle-relapsex

  leakage-aware, causality-aware relapse prediction for systemic lupus erythematosus — a different target from slecare, and a stricter evidence bar.

  0.667f1 · catboost

  silver awardco-author · rich softam 2026 symposium · 2026

  • contributed to a relapse-prediction framework over a stratified 120-patient retrospective cohort, designed so no feature could leak the outcome it was meant to predict.
  • helped extend standard shap with correlation-aware conditional shap and exploratory double machine learning, isolating a clinically consistent haemoglobin threshold near 11.75–11.95 associated with elevated relapse risk.

  conditional shapdouble mlleakage analysis

• cinemetric

  do actor facial expressions predict how an audience receives a short film? the honest answer turned out to be no.

  nulland reported

  null result, published as onefinal-year project · ftsm ukm · 2026

  • built an end-to-end pipeline of yolov5-face detection, cnn-resnet emotion feature extraction, and xgboost regression across films running 3 to 45 minutes.
  • an optimistic single train-test split looked publishable; repeated k-fold cross-validation collapsed it. replaced the split, re-ran the full evaluation, and reported the null result.
  • prevented a false-positive finding from reaching publication — the finding held under repeated cross-validation.

  yolov5-faceresnetxgboostcross-validation

• goguard

  graph-based scam detection that flags fraudulent recipients before the transfer clears, not after the report comes in.

  0.8621precision

  selected for publication with ukmtngdigital fintech hackathon · jan 2026

  • led ml development on a detection pipeline pairing graph neural networks over phone-number transaction topology with an xgboost risk classifier.
  • engineered network-topology features exposing coordinated recipient clusters invisible to per-transaction scoring, hitting 0.8621 precision and 0.7901 recall on held-out data.
  • the architecture was selected for continued academic publication with ukm.

  graph neural networksxgboostfraud topology

• resumeje

  an ai-assisted resume platform built around achievement banking and target-specific tailoring for internships, scholarships, and fellowships.

  silver awardfirst author · rich softam 2026 symposium · jan 2026

  • combined a next.js and fastapi system with ollama-based ai tailoring, a rules-based fallback for when the model is unavailable, and a deterministic rendercv export pipeline.
  • delivered an auditable admin review workflow over immutable resume snapshots, supporting reviewer comments, resubmission cycles, and notifications; validated with functional use-case testing and automated backend tests.

  next.jsfastapiollamarendercv

• kdstudio

  ai-enabled staff management for a hybrid, distributed content organisation — attendance, location, tasks, and payroll in one android app.

  15usability testers

  gold awardco-author · rich softam 2026 symposium · 2026

  • co-developed a platform integrating facial-recognition attendance, gps location verification, task coordination, and automated payroll generation on jetpack compose, firebase/firestore, and google ml kit.
  • contributed an ai virtual assistant for organisational queries using gemini ai and tensorflow lite, validated through black-box functional testing and usability feedback from 15 internal users.

  jetpack composefirebaseml kittensorflow lite

• brugada syndrome classification

  cardiac risk classification on the official physionet brugada-huca dataset, tuned rather than guessed.

  0.9494roc-auc

  international data science competition (idsc) malaysia · jan 2026

  • classified brugada syndrome cases at 0.8949 f1 and 0.8904 accuracy alongside the 0.9494 roc-auc, tuning an xgboost classifier through optuna-driven hyperparameter search under a stratified validation scheme.

  xgboostoptunaphysionet

• national youth roundtable 2026 report

  digital diplomacy, the us–iran war, and youth-led multilateral reform — a national policy report that left the campus.

  3policy domains

  presented to mofa, the un, and aseanco-author · malaysian youth diplomacy · 2026

  • authored domain 2 of the three-domain report: ceasefire mediation architecture, malaysia's energy and trade exposure through the strait of hormuz, and the enforcement gap between international law and live conflict.
  • co-authored the concluding synthesis, distilling cross-domain findings into a sequenced institutional reform agenda spanning digital resilience, ai accountability protocols, and a statutory youth foreign policy council.

  ai governancepolicy synthesismultilateral reform

• routeiq

  an ai sales copilot that routes field visits by expected customer value rather than travel distance.

  1stof 20 teams

  1st placegenerative ai vibe hackathon, ftsm ukm · jun 2026

  • engineered an xgboost scoring pipeline ranking customers by deal potential, rendering each score as a plain-language visit recommendation by layering shap-style explainability and llm reasoning over the model output.
  • exposed live route plans, customer scores, and visit recaps to external ai agents including chatgpt and claude through a model context protocol access layer.

  react pwafastapisqlitemcpxgboost

• clover

  explainable cloud security: detect, recommend, heal — with the reasoning grounded in real telemetry.

  4thof 50 teams

  4th placeimaginehack, taylor's university · jun 2026

  • designed a pipeline prioritising cloud infrastructure vulnerabilities using live telemetry and cvss severity scoring.
  • cut operator triage time on ambiguous alerts by grounding llm root-cause analysis and remediation recommendations in a vector-database rag layer built over cloud telemetry and operational data.

  ragvector dbcvsscloud telemetry

• local-first multilingual rag engine

  retrieval-augmented generation with no api key, no per-query cost, and no english-centric assumptions.

  0external api calls

  independent research project · 2026

  • eliminated all external api dependency and per-query inference cost by implementing a fully local, llm-free engine in pure numpy with a hierarchical tree-based hybrid retrieval strategy.
  • extended reliable retrieval to bahasa melayu alongside english by designing multilingual embedding and scoring paths that avoid english-centric tokenisation assumptions.

  numpyhybrid retrievalbahasa melayuembeddings

• online safety irl

  carried malaysian ai governance and anti-scam policy from a campaign into ambassadorial and parliamentary rooms.

  malaysia representativeyouths irl fellow · oct 2025 — apr 2026

  • represented malaysia at the asean youth leaders summit 2025, presenting ai governance and anti-scam policy positions to eu ambassadors, technology leaders, and government officials.
  • extended the campaign to a national broadcast audience on tv9 and exchanged policy papers with the ministry of communications at the malaysian parliament, translating technical scam-detection concepts into public-facing policy language.

• malaysian youth sdg summit 2026

  delivered the ai, skills, and the future of work panel alongside a talentcorp representative.

  panellist · jul 2026

  • framed five-year workforce trends, ai employability risk, and concrete skill pathways for new entrants to the job market.

• school@ukm arvis robotics workshop

  a nationwide stem robotics workshop and competition, run end to end.

  rm16kclub revenue

  programme lead · sep 2024 — jan 2025

  • led nationwide school recruitment and ran the workshop and competition end to end, generating rm16,000 in club revenue and the highest participation on record for ukm's robotics club.

• jpmpp ukm

  a non-partisan, university-wide civic awareness campaign, measured in turnout.

  67.35%voter turnout

  head of political education · oct — dec 2024

  • designed and executed the campaign that drove campus turnout to 67.35% — the highest recorded in ukm history.

• tedxukm

  sponsorship pipeline and speaker sourcing across two flagship events.

  5partnerships secured

  head of partnership · dec 2023 — apr 2025

  • secured five corporate and association partnerships across two consecutive flagship events by building a sponsorship pipeline and negotiating value-aligned packages.
  • brought award-winning rural stem educator anuthra sirisena to the 2024 flagship stage through speaker sourcing and outreach.

languages

pythonsqljavascript / typescripthtml / css

machine learning

xgboostcatboostrandom forestsvmmlpgraph neural networksensemble stackingscikit-learnpytorchoptunashapcross-validation design

llm & nlp

fine-tuningragvector databasesembeddingsqwenollamamodel context protocolsynthetic qa generationprompt engineering

computer vision

yolov5-facecnn / resnet feature extractionfacial expression recognition

data engineering

pandasnumpypymupdfetl & ingestion pipelinesedacleaning & deduplicationweb scraping

backend & deploy

fastapinext.jsreact (pwa)sqliterest apisvercelgit

spoken

bahasa melayu (native)english (professional working)

education & credentials