I ship detection systems that catch abuse in production.
Focus on phishing, fraud, and malicious content. I work across CV + NLP + SecML — from data to ONNX API with Docker. Production-first, not research demos.
Detection systems, shipped to production.
Phishing & Malicious URL Detection
NLP models that flag phishing and malicious URLs in real time. From curated datasets and ensemble training to low-latency ONNX APIs.
Fraud & Anomaly Detection
Detectors for fraudulent transactions and anomalous behavior. Tuned for low false positives on production traffic.
Computer Vision Safety
Real-time vision for safety and abuse detection. Dataset pipelines, CNN training, and edge inference with OpenCV.
Projects that shipped to real users.
ML and security systems focused on detection, low false positives, and production deployment.
PhishGuard: Threat Detection API
Security-focused ML API for phishing detection optimized with ONNX.
Existing phishing APIs were too slow for real-time edge deployment and suffered from high false positives.
Trained a Custom NLP text classifier using Scikit-Learn. Converted the model to ONNX runtime for sub-millisecond inference and containerized with Docker and Flask.
Achieved 99.2% accuracy in phishing detection with a 40% reduction in inference latency compared to traditional REST setups.
Zero-Day Malicious URL Scanner
Threat intelligence platform categorizing web links using ensemble classifiers.
Attackers frequently mutate URLs to bypass signature-based blocking.
Engineered an ensemble model (Random Forest + XGBoost + SVM) to classify URLs based on lexical features combined with real-time WHOIS heuristics.
Effectively caught 92% of zero-day morphed malicious links with a false positive rate of < 1.5%.
Alcohol Detection CV System
Hybrid AI system for real-time alcohol intoxication detection using facial cues.
Needed a non-invasive way to detect intoxication for safety integrations without breathalyzers.
Curated dataset of ~100k facial images. Developed a hybrid system using CNNs for feature extraction and SVM for classification, deployed as an Android app.
Delivered a lightweight mobile model capable of 15fps inference on mid-range Android devices with strong predictive reliability.
AI Fraud Detection System
Classification pipeline engineered to detect fraudulent transactions.
High false-positive rates in transaction alerts caused customer abrasion and review fatigue.
Implemented cost-sensitive learning algorithms and SMOTE for imbalanced classes, prioritizing Precision and minimizing False Positives over standard Accuracy.
Cut false positives by 65% while maintaining recall, saving significant manual review hours.
Tech behind the systems.
AI engineering + security + production deployment. No bloat, just tools that ship.
Proficient
Security Stack
Deployment
From full-stack to ML systems.
AI/ML Engineer
Building ML pipelines for threat detection, phishing, and CV. From dataset engineering to lightweight ONNX APIs deployed in production. Focus on low FPR and real-time inference.
Software Engineer — Full-Stack
End-to-end ownership across web and mobile. Built scalable, high-performance systems with modern frameworks and cloud infra. Full SDLC from architecture to deploy.
Services built to ship in weeks.
Targeted AI/ML engineering focused on production-ready defense mechanisms, not demos.
Threat Detection API Development
Integrating ML models natively into backends is slow, bloated, and hard to scale under high throughput.
Lightweight, Dockerized inference APIs using ONNX runtime and Flask. Optimized for <15ms latency.
A standalone, highly available endpoint for real-time safety classification.
CV Safety System Deployment
Many CV proof-of-concepts fail to transition into stable, edge-deployed tools.
From dataset curation to robust CNN pipelines, delivering production-ready vision that handles dynamic environments.
A reliable vision engine ready for your mobile or web apps.
Security Audit for Existing ML
Classification pipelines suffer from high False Positive Rates, causing review fatigue.
Analyze dataset, model bounds, and retrain with cost-sensitive learning focused on precision.
Tightened precision, fewer false alerts, protected core ops.
Let's build a detection system that ships.
Based in PHT (GMT+8), remote-first. I take on ML security roles and freelance contracts for threat detection, CV safety, and precision tuning.