Cloud engineer, 3+ years on Google Cloud and AWS — mostly VPCs, GKE, and state files. Previously at an MNC. I terraform plan more times than strictly necessary.
Right now I'm trying to teach a Kubernetes scheduler to care how much power it's burning.
- Federated Reinforcement Learning for energy-efficient cloud data centres — privacy-preserving RL scheduling across heterogeneous GKE clusters. PPO with FedProx + SCAFFOLD, a pointer-style masked actor-critic, and a Kubernetes scheduler extender running in shadow mode. Energy measured against real hardware (RAPL + a metering PDU), because the software estimator lies and I can prove it. Benchmarked against centralized-RL and local-only-RL bounds.
- A Clash Royale stats pipeline — Wilson score intervals on my own losses, segmented by mode, archetype, trophy band, and season. Time-aware, because lifetime averages flatter you. It has not made me better at the game. Or did it??
- A Spotify library manager — for someone who refuses to sort playlists by hand. Spotipy against the Web API, with track characteristics pulled from ReccoBeats because Spotify deprecated
/audio-featuresout from under everyone. Three years of Last.fm scrobbles reconciled in via pylast, landed in SQLite, surfaced through a local Streamlit UI.
Day job is the boring half of the stack: networking, IAM, and whatever the scheduler decided to do at 3am. Most of the above lives in private repos for now.
Collector of Google Cloud certifications, reluctant participant in cost optimization meetings. All verifiable on my Google Developer Profile.
