Data that drives
decisions.
Data Analyst & Engineer building production-grade systems, grounded in AI/ML engineering to solve real-world problems.
I build AI systems that are production-ready — not just demo-ready.
About
From Jakarta to Singapore — building production systems along the way.
I'm a Computer Science (Big Data) student at University of Wollongong in Singapore, focused on building AI systems that are production-ready — not just demo-ready.
As Team Lead, I built SecureAdvisor for Certis Group — an AI-powered security incident management platform that fuses live CCTV, access logs, and manual triggers using YOLOv8 detection and GPT-4o advisory to coordinate real-time ground officer dispatch across a 3-app system.
In an industry collaboration with Aires Applied Technology, I built MakanMap — a real-time crowd forecasting system powered by a Gradient Boosting Regressor (93.2% accuracy), served through a FastAPI backend and a React dashboard with live What-If scenario analysis.
At SIM Data Analytics Club, I developed a Real Estate Valuation Analyzer trained on 1M+ residential records — covering end-to-end ML inference, real-time over/undervaluation analysis, and a recommendation engine that cut property search time by 15%.
What drives me is turning messy real-world problems into systems that actually work — clean pipelines, reliable APIs, and interfaces that make decisions easier for the people using them.
What drives me
Data systems that scale. From raw pipeline to production interface — reliable, observable, and engineered to turn big data into real business impact.
Current obsession
AI/ML Pipeline Design — the infrastructure layer that makes data systems actually work in production. Observable, scalable, and debuggable.
University of Wollongong
Bachelor of Computer Science (Big Data)
Sep 2024 – Dec 2026 · Singapore
Singapore Institute of Management
Diploma in Information Technology
Oct 2023 – Sep 2024
Experience
Internships, Industry Projects, and Clubs.
Frontend Engineer
ContractAires Applied Quantum Technology
Jan 2026 – Mar 2026 · Singapore · Remote
Built MakanMap — a real-time crowd level monitoring dashboard in React and Vite, surfacing ML predictions at 93.2% classification accuracy across 30-minute forecast bins up to 3 hours ahead. Built a What-If scenario analysis module for instant side-by-side comparison of crowdedness predictions. Containerised a 4-service stack with Docker Compose and automated CI/CD via GitHub Actions.
Data Analyst
ContractSIM Data Analytics Club
Sep 2025 – Mar 2026 · Singapore · On-site
Developed a real estate valuation analyzer platform — a Gen-AI property valuation system trained on 1M+ residential records covering end-to-end ML inference, real-time over/undervaluation analysis, and a recommendation engine that cut property search time by 15%.
Software and Web Developer Intern
InternshipAstrindo Senayasa
Apr 2025 – Jun 2025 · Jakarta, Indonesia · On-site
Developed an AI chatbot for natural language product search and multi-product comparisons across 100+ IT products. Implemented automated report generation — reducing database load by 45%. Enhanced internal query resolution time by 35% through contextual memory tracking.
Data and IT Committee
ContractPPI Singapura
Jan 2025 – Nov 2025 · Singapore · On-site
Led a cross-functional team of 5+ to deliver a university comparison platform for 10+ institutions. Drove a 20% increase in page engagement through UI restructuring. Managed 50+ digital assets as part of PPI Singapura's digital transformation initiative.
Selected Work
SecureAdvisor
Built for Certis Group — this platform coordinates real-time security incident response across a 3-app system, led as Team Lead with a cross-functional team.
On the detection side, CCTV frames, access control door events, and manual panic triggers are routed through YOLOv8n at 0.7 confidence with per-camera polygon zone validation.
A sliding event window de-duplicates and fuses signals across all three streams, classifying them into 7 incident types before escalating to GPT-4o — which returns a threat severity flag, a structured response plan, and a named dispatch unit with ETA.
What makes this distinct is the end-to-end pipeline latency under 2 seconds — from raw CCTV frame to actionable officer dispatch recommendation — running across a FastAPI backend and 3 React + Vite frontends.
System Architecture
< 2s
End-to-end pipeline — from raw CCTV frame to YOLOv8n detection, rule correlation, and GPT-4o advisory with officer recommendation
0.7
YOLOv8n confidence threshold for real-time person detection with per-camera restricted zone polygon checks
7+
Incident types detected by the rule-based engine — intrusion, loitering, tailgating, panic, fire, unauthorized access, after-hours presence
120s
Sliding event window for multi-source signal fusion — correlating CCTV, access logs, and manual triggers with 30s duplicate suppression
MakanMap
An industry collaboration with Aires Applied Technology, a Singapore-based deep-tech startup — built to give venues and users a reliable way to plan around crowd density before it happens.
The forecasting engine is a Gradient Boosting Regressor trained to 93.2% accuracy on contextual signals: temperature, humidity, weather condition, time-of-day, public holiday flags, and historical location frequency. Data flows through an Apache Airflow pipeline — two sequential DAG tasks handle raw event cleaning and feature engineering, writing directly into a Supabase PostgreSQL feature store before predictions are served via FastAPI.
The dashboard goes beyond a chart — a What-If scenario analysis module lets users override input conditions and compare alternate crowd outcomes side-by-side against the baseline. The entire stack runs across 4 Docker services with GitHub Actions CI/CD, making it deployable and reproducible out of the box.
System Architecture
93.2%
Gradient Boosting Regressor accuracy across 30-minute crowd level bins — trained on weather, time, location frequency, and public holiday signals
3 hrs
Max forecast horizon in the dashboard — up to 6 × 30-minute bins with Low, Medium, and High crowd classification
4
Docker services via Compose — FastAPI backend, React dashboard, MLflow experiment tracker, and Airflow pipeline scheduler
2
Airflow DAG tasks on schedule — raw data cleaning then feature engineering writing directly to Supabase's features table
Real Estate Valuation AI
A property valuation tool built for the SIM Data Analytics Club, trained on ~1M Zillow residential records spanning all 50 US states — the system takes a listing URL and outputs a verdict.
The ML pipeline tackles a tricky data problem — city and state are high-cardinality categoricals that break naive encoding. The solution uses K-Fold cross-validated Target Encoding, replacing each label with a target-mean value to preserve strong location price signal without data leakage. XGBoost Regressor and Gradient Boosting Regressor were benchmarked via GridSearchCV, with the best model serialised as a full Scikit-learn Pipeline.
The output isn't just a number. A Bullet Chart compares predicted vs actual price to surface the over/undervaluation percentage, while a GPT-4o advisory layer generates conversational valuation context and 4% annual forward price projections — all accessible through a Gradio web interface.
System Architecture
~1M
Zillow residential records across all 50 US states — covering price, city, state, lot size, house size, beds, and baths for training
15%
Property search time reduced via AI recommendation engine — surfacing over/undervalued listings ranked by valuation delta
2
Ensemble models benchmarked with GridSearchCV hyperparameter tuning — XGBoost Regressor vs Gradient Boosting Regressor
4%
Annual growth rate applied for forward price projections from 2025 — embedded in the GenAI valuation advisory output
Astrindo Chatbot
Built during my internship at Astrindo Senayasa (Jakarta) — an internal enterprise chatbot that lets non-technical employees query live business data across Marketing, HR, Finance, Purchasing, and Service departments using plain language.
The system runs a two-stage NLU pipeline: GPT-4o-mini classifies the user's intent and extracts structured entities (year, month, specialist name, city) at temperature=0, returning strict JSON. The PHP backend then dispatches to a domain-specific feature handler that executes deterministic MySQL queries — the LLM never touches the database, making hallucinated figures structurally impossible.
The NLU call also generates a ChatGPT-style sidebar title per session, with a hard-banned list of generic labels and a regex-based Indonesian fallback to keep titles specific and natural across both English and Indonesian input.
System Architecture
12
Intents across 5 business domains — Marketing, HR, Finance, Purchasing, and Service — each with a dedicated PHP feature handler
0
Hallucinated numbers — LLM returns only intent + entities; all figures come from deterministic MySQL queries against live business data
2
Languages supported — bilingual input (Indonesian and English) with intent detection stable across both via the same NLU prompt
5
Business departments integrated — Marketing, HR, Finance, Purchasing, Service — covering the full Digital Approval reporting scope
Certifications
Verified credentials from IBM and industry platforms.