real-time-ml

Real-time fraud detection system using ensemble ML models, featuring streaming data processing, explainable AI with SHAP, and production-ready deployment with FastAPI and Docker.

A source-available JVM pipeline kernel for policy-aware, benchmarkable operational data movement.

Real-time UPI fraud detection system (0.8953 ROC-AUC) with <500ms FastAPI scoring, 480+ temporal features, and budget-aware alerts under fintech constraints

Production-grade fraud detection pipeline with entity-level behavioral feature engineering, velocity anomaly detection, graph-based risk signals, and real-time scoring API. Built with XGBoost, PyTorch, FastAPI, and Docker.

Real-time AI-enabled hand gesture recognition system for deaf and mute communication using OpenCV, MediaPipe, and Python

Production-style real-time ML feature store with low-latency inference

Real-time anomaly detection pipeline — Kafka → FastAPI → Isolation Forest + Autoencoder → WebSocket dashboard. Sub-500ms end-to-end latency.

Low-latency financial fraud detection engine utilizing LightGBM and Redis in-memory feature caching, evaluated strictly on PR-AUC and Precision@Top-100 for extreme class imbalance.

Real-time recommendation platform: exactly-once Kafka sinks to Iceberg and Delta on Flink, Feast feature store (Iceberg/Delta offline, Redis online), FastAPI/Redis serving, Chaos Mesh on GKE.

I built an end-to-end ML system that takes a transaction, engineers real-time behavioral features, generates a fraud-risk prediction through an API, stores and monitors the prediction, visualizes the results on a dashboard, and uses automated testing and CI/CD to maintain reliable deployment.

Event-driven stream processing for real-time ML inference.

Código fuente: Análisis de Vuelos basado en trabajo de Valliappa Lakshmanan.

Real-time fraud detection platform with automated MLOps pipeline — experiment tracking (MLflow), data drift monitoring (Evidently), FastAPI serving, Docker, CI/CD via GitHub Actions, and cloud deployment.

Real-time fraud detection engine - Kafka → Spark → dual-write Redis/Snowflake feature store → XGBoost, served via FastAPI in <50ms. Catch fraud in milliseconds. Streaming feature store (Kafka + Spark) with zero train/serve skew, served live via FastAPI.

2nd place 🏆, Precision NeuroScience's BCI Hackathon 2026 — streaming neural decoder that compresses 1024 neural channels to 64 in real time using FFT spectral features and a sub-25MB PyTorch MLP with microsecond inference latency.

Production-grade ML platform portfolio for real-time decisioning systems.

M.S. thesis: real-time MUSE 2 EEG+PPG stress detection and severity classification integrated with SeeBT for closed-loop intervention.

Real-time transaction fraud risk scoring for Acquirer clients — Straive Strategic Consulting

Client-server backend for real-time Activity of Daily Living (ADL) classification. Handles window ingestion, prediction, label requests, and incremental model updates.