HYPERION: A production-grade Multi-Agent Reinforcement Learning (MARL) system that coordinates autonomous UAV swarms for hypersonic threat detection, tracking and interception in high-speed, GPS-denied environments. Built for aerospace and defense applications, this project showcases MLOps best practices, distributed systems design, and advanced reinforcement learning techniques applicable to autonomous systems challenges.

Key Features

• Multi-Agent Reinforcement Learning: Implemented MAPPO (Multi-Agent Proximal Policy Optimization) to enable 20+ autonomous agents to coordinate interception strategies without centralized control. Graph Neural Networks handle dynamic agent communication patterns.
• Physics-Informed Simulation: Integrated Physics-Informed Neural Networks (PINNs) to generate realistic hypersonic trajectories (Mach 5+) with accurate aerodynamic modeling, providing high-fidelity training environments.
• Sensor Fusion Architecture: Combines RF positioning (TDOA/FDOA), thermal imaging, and telemetry streams using Extended Kalman Filtering for robust threat tracking in contested environments.
• Production-Grade Implementation Scalable Training Pipeline | Containerized Deployment | Comprehensive Testing | Interactive Dashboard

SENTINEL: A production-grade Sensor Fusion system combining Overhead Persistent Infrared (OPIR) thermal detection with Radio Frequency (RF) geolocation for real-time threat detection and tracking. Implementing advanced signal processing, SENTINEL showcases multi-sensor fusion fundamentals, signal processing expertise, and multi-modal data integration relevant to intelligence and defense applications.

Key Features

• Multi-Sensor Fusion Architecture: Implemented Extended Kalman Filtering to fuse asynchronous data streams from RF positioning (TDOA/FDOA), thermal imaging arrays, and telemetry feeds. Handles varied sensor hardware profiles with robust error handling for missing or degraded signals.
• RF Positioning System: Developed Time Difference of Arrival (TDOA) and Frequency Difference of Arrival (FDOA) algorithms for passive emitter location. Achieves high-accuracy triangulation using multiple receiver stations without requiring active transmission.
• Real-Time Processing Pipeline: Architected asynchronous data ingestion supporting high-frequency sensor streams (100+ Hz thermal, intermittent RF bursts) with sub-second latency for threat detection and track updates.

MERIDIAN: A Quantitative Portfolio Optimization Engine implementing modern portfolio theory for dynamic asset allocation and risk management. Built for quantitative investment management, this project showcases financial engineering expertise, optimization techniques, and data engineering practices relevant to fintech, asset management, and proprietary trading firms.

Key Features

• Multi-Strategy Optimization: Implemented Mean-Variance Optimization (MVO) and Risk Parity allocation strategies using CVXPY for convex optimization. Solves for optimal asset weights under complex constraint specifications including sector limits, turnover costs, and regulatory requirements.
• Real-Time Market Integration: Engineered data ingestion pipeline connecting Yahoo Finance API to stream live market data, pricing feeds, and corporate actions. Processes multi-asset class data (equities, fixed income, commodities) into standardized formats for model consumption.
• Advanced Risk Modeling: Built correlation analysis framework incorporating rolling windows, exponential weighting, and regime detection to capture time-varying relationships between assets. Generates dynamic covariance matrices for accurate risk quantification across market conditions.
• Dynamic Rebalancing Logic: Developed trigger-based rebalancing algorithms that balance transaction costs against drift from target allocations. Incorporates tax-loss harvesting opportunities and liquidity constraints for practical implementation.

CONSTELLATION: A production-grade predictive maintenance platform for Low Earth Orbit (LEO) satellite constellation monitoring. Combines survival analysis, anomaly detection, and degradation forecasting to predict subsystem failures and optimize maintenance scheduling. Built for space operations management, this project showcases time-series forecasting, survival analysis, cloud-native MLOps, and predictive maintenance techniques relevant to aerospace, defense, critical infrastructure management, and industrial IoT applications.

Key Features

• Multi-Subsystem Health Monitoring: Engineered predictive models for four critical satellite systems: power (battery degradation, solar panel efficiency), thermal (component temperature anomalies, radiator performance), attitude control (reaction wheel bearing wear), and communications (transponder degradation, antenna drift).
• Survival Analysis Framework: Implemented Cox proportional hazards and Weibull analysis for time-to-failure prediction, providing probabilistic failure forecasts with confidence intervals. Enables proactive maintenance scheduling before critical component loss.
• Advanced Anomaly Detection: Deployed Isolation Forest and Autoencoder models to detect unusual telemetry patterns in real-time, flagging deviations from operational baselines before cascading failures occur. Handles high-dimensional sensor data across multiple satellites simultaneously.
• Degradation Forecasting Pipeline: Built LSTM and Temporal Fusion Transformer models to predict performance decline trajectories for solar panels, batteries, and mechanical systems. Captures long-term trends while accounting for orbital environment effects (thermal cycling, radiation exposure).

Data Science | ML Engineering