ML / MLOps Engineer · IEEE-published researcher · AWS Certified · M.S. student at ASU
Building GPU training reliability, telemetry pipelines, and secure ML in applied aerospace and industrial systems.

📍 Phoenix, Arizona, US
🔗 LinkedIn · 💭 Open to work

I build production-minded ML systems for telemetry-heavy, safety-conscious environments, with a focus on GPU reliability, anomaly detection, forecasting, and secure deployment. My work sits at the intersection of machine learning, MLOps, and aerospace-style monitoring systems.

I’m especially interested in:

• ML infrastructure and model reliability
• Telemetry pipelines and anomaly detection
• Applied aerospace and mission-control systems
• Secure, auditable, Dockerized ML deployment

What: Prototype for modeling GPU memory fragmentation and reducing training-time out-of-memory failures on RTX-class workloads.
Why: Long-running deep learning jobs fail expensively when fragmentation causes avoidable OOM crashes.
How: Built a PyTorch- and Docker-based experimental setup to evaluate fragmentation-aware mitigation strategies on GPU training workloads.
Metrics: OOMs reduced 23 → 0, VRAM utilization 94% → 87%, fragmentation ratio 0.61 → 0.18, training overhead < 2%.

What: Predictive maintenance pipeline using NASA C-MAPSS data for remaining useful life forecasting and anomaly detection.
Why: Industrial and aerospace maintenance systems need earlier fault visibility and more accurate degradation forecasting.
How: Designed an MLOps-oriented pipeline with Dockerized local deployment, time-series modeling, and telemetry-driven anomaly workflows.
Metrics: RUL RMSE 166.7 (~10% over linear baseline), anomaly F1 0.373, throughput 52,368 req/s, P95 latency 3.94 ms, 52 automated tests.

What: Hybrid orbital monitoring project combining ESP32 sensing, Kalman filtering, re-entry simulation, and a containerized Streamlit dashboard.
Why: Telemetry systems are more useful when hardware, simulation, filtering, and operator visibility are connected end to end.
How: Integrated IoT sensing, state estimation, simulation, and dashboard-based monitoring into one reproducible system.
Metrics: Telemetry streaming at 10–50 Hz per device, Kalman filter reduces variance by ~50%, re-entry simulation from 550 km LEO to impact.

What: Physics-aware aerospace trajectory simulation sandbox using RK4 integration, atmospheric models, and ML surrogates.
Why: High-fidelity simulation is useful for testing constrained inference and trajectory behavior before deployment.
How: Combined classical numerical methods with ML surrogate modeling in a Dockerized Python environment.
Metrics: Surrogate speedup 11.5x over RK4 (P99 latency 1.84 ms → 0.16 ms), throughput 500 → 5,750 seq/s, bit-accurate parity (MAE 0.0000 m).

What: ESG telemetry and carbon analytics platform for async pipelines, forecasting, anomaly detection, and audit-friendly reporting.
Why: Sustainability reporting requires traceable data pipelines and reliable anomaly surfacing, not just dashboards.
How: Built an async telemetry-focused architecture with Python, Docker, PostgreSQL, and forecasting workflows.
Metrics: Ingestion latency p99 450 ms → 42 ms (~10.7x), forecast MAE 4.2% (vs 14.2% baseline), anomaly recall 94.2%, Merkle audit 13.6x faster.

What: Mission-control telemetry simulation and anomaly surfacing stack for orbital monitoring and operator-facing dashboards.
Why: Operators need interpretable health signals and real-time visibility into changing system states.
How: Built telemetry simulation flows and Streamlit-based monitoring views for anomaly detection and operational review.
Metrics: EKF state convergence in ~10 steps, GA delta-v optimization < 1s, anomaly inference < 50 ms per telemetry frame, full Space-Track TLE catalog (~4.6 MB) rendered live.

What: CubeSat telemetry monitoring pipeline for anomaly detection, Firebase-backed data flow, and satellite health analytics.
Why: Small-satellite systems need lightweight but reliable telemetry infrastructure for health assessment.
How: Combined telemetry ingestion, cloud-backed data flow, and anomaly analytics into a deployable monitoring pipeline.
Metrics: Ensemble F1 0.928, precision 0.942, recall 0.915, throughput 63,622 events/s, inference latency 15.72 µs, false alarm rate ~3–5% over 24h window.

Languages & ML: Python, PyTorch, time-series modeling, anomaly detection, forecasting
MLOps & Infra: Docker, CI-friendly project structure, reproducible local deployment, telemetry pipelines
Data & Systems: PostgreSQL, Firebase, IoT telemetry, monitoring dashboards
Domains: Aerospace systems, predictive maintenance, ESG analytics, mission-control simulation

• Building ML systems that survive real operational constraints
• Turning telemetry into actionable health and reliability signals
• Combining simulation, monitoring, and inference in one workflow
• Writing clean, reproducible, recruiter-readable project documentation

• GPU training reliability and fragmentation-aware optimization
• Production-ready ML observability
• Secure ML deployment patterns
• Applied robotics and aerospace intelligence systems