About
Built by SREs who got woken up at 3am too many times.
Crest exists because drive failures were still surprising infrastructure teams in 2024. Dozens of storage telemetry signals were sitting unused inside every piece of hardware. We built the model to read them.
The problem we kept running into.
S.M.A.R.T. data has been available on hard drives since the mid-1990s. By 2024 every NVMe and SSD shipping into data centres exposed dozens of wear and health attributes. Most teams had monitoring set up, but it only showed raw values without any context for what a trend meant or when it was about to matter.
The teams that knew what to look for, watched individual attribute time-series in Grafana and built custom threshold rules. It worked, but it required institutional knowledge to maintain and still produced a steady stream of false positives. Every real failure was preceded by days of signal that no one had time to correlate.
We found that a survival model trained on attribute trajectories outperformed threshold alerting on every failure class, across multiple drive generations and vendors. Crest is that model packaged as an agent that fits into the tooling infrastructure teams already run.
The team.
Led infrastructure SRE at a high-growth fintech for four years before founding Crest. Rebuilt their on-call system after two data-loss incidents tied to undetected NVMe degradation. Computer Science, IIT Bombay.
Research background in reliability engineering and survival analysis before moving into applied ML. Built the first version of the failure prediction model in 2023 as a side project after losing a personal NAS to sudden SSD failure.
Previously at Zepto and Razorpay working on distributed storage and observability infrastructure. Responsible for the Crest agent, ingestion pipeline, and on-premise deployment stack. Deep knowledge of Linux storage subsystems.
Work with us.
We are a small team and stay that way intentionally. If you want to work on storage reliability infrastructure at an early stage, email us. We are not hiring broadly, but we talk to people who are obsessive about the problem.
Email [email protected]Get early access.
Crest is in private early access. We are onboarding infrastructure teams who are actively managing storage fleets and want to move from reactive replacement to predictive swap scheduling.
Apply for Early Access