projects

Production ML systems and personal builds — every metric on these cards is a real, verified number.

▲ 45–55% runtime cut

Reinforcement-learning service for self-optimizing 5G/6G networks. I built the MRO module end-to-end and re-architected the training pipeline from linear to multi-threaded, cutting runtime by 45–55%. Unified RL harness across MRO, CCO, Energy Saving, and Load Balancing rApps.

  • python
  • pytorch
  • ppo
  • o-ran
◆ 3GPP-compliant

Golden topology generator powering Maveric's rApps. I wrote the 3GPP-compliant golden generators — Hata/UMa/UMi path loss, Gauss-Markov mobility across four UE velocity classes, and ECI computation.

  • python
  • numpy
  • 3gpp
  • kafka
◆ 3-tier cache

GPyTorch Gaussian-process engine for RF propagation prediction. I hardened the serving layer: 3-tier cache (Redis → MongoDB → Postgres), partition-aware exact-offset Kafka commits, and Prometheus-instrumented training histograms.

  • python
  • gpytorch
  • fastapi
  • redis

MalwareSentinel

▲ adaptive RL firewall

RL-based adaptive firewall with real-time malware detection. Learns from network activity and shares collaborative threat intelligence across devices — ML models and cloud integration designed for fast, device-specific protection.

  • python
  • rl
  • security
◆ 7-class ResNet

Seven-class facial-emotion classifier fine-tuned on FER2013 — ResNet backbone with class-weighted loss for imbalance, test-time augmentation, and multi-GPU training over NCCL.

  • python
  • pytorch
  • resnet
▲ zero dependencies

Zero-dependency PWA for volleyball referees and scorekeepers. Tracks scores, rotations, substitutions, timeouts, and libero usage in real time with automatic rule enforcement. Vanilla JavaScript.

  • javascript
  • pwa