The setup
RMS had a reputation as a firm full of top-flight professionals, Stanford PhDs and the like, and was the name in complex big-data catastrophe computation. Their product was a risk-modeling system built around exceedance-probability curves, or EP curves, the core output insurers and reinsurers use to price and manage catastrophe risk. At the time, it shipped as a desktop application bundled with a precomputed set of models as a binary. RMS hired me as an engineer as the company was moving to re-platform the whole thing as SaaS, shifting from shipping a desktop binary to delivering the models and computation as a hosted service.
My role
Hired as an engineer on the SaaS re-platforming, moving the catastrophe-risk modeling capability from a shipped desktop binary toward hosted, service-delivered computation. It was a Microsoft-heavy shop, so a natural fit. I got coding immediately and quickly became the architect for the cloud-hosted initiative that Craig Schultz was leading. He was then ~VP of Product Management, and came from Lawrence Livermore National Lab. Craig and I got on well and spent a lot of time on RMS’s hard problems together.
What I built
The hard part was turning catastrophe computation into a secure service. RMS models push very large data models through heavy computation, so moving to SaaS meant answering questions the desktop binary never had to. How do you offer it as a service, securely? How do you price running those computations on hosted infrastructure? And how do you absorb surge demand from a central platform, the “a hurricane is coming, recompute the models now” request, with core services split across C# and Java?
My focus was the compute-scheduling and memory layer. That meant scheduling and prioritizing many processing requests across multiple models at once, loading models on demand and unloading them to fit memory, and keeping it efficient on a scaled, shared architecture. The work also covered the usage and pricing model for that hosted computation, all with the security the data and models demanded. Hard, interesting systems work.
Net result
RMS comes down to three pieces. First, the joy of joining a world-famous organization. It was one of the rare times I applied blind and landed somewhere that dense with top-flight talent. Second, an interesting architecture challenge: shared, hosted processing of complex data models and heavy mathematics, offered securely as a service, with the scheduling, memory, and pricing problems that came with it. And third, the cue to leave. A reorg pushed the platform onto a Java stack, moving the work away from my core competency, C# and .NET and the systems layer I do best. I’d done real systems architecture and built a strong reputation fast, and then it was time to go.
The story
I was elated to join RMS. It was one of the few times in my career I applied for a job blind, not through my network, and landing it at a firm that dense with top-flight talent meant a lot.
Then came a wild reorganization. RMS brought in a new engineering leader and, with him, much of his former development org, and reorged me under him. I’d built a strong reputation fast through the architecture work, but it wasn’t a fit. It was now a Java shop, so I gave notice and left. My elation at joining ended faster than I’d have liked. Still, I’d done real systems architecture.
What I learned
- Catastrophe computation as a secure service is a real systems discipline. It takes scheduling, memory-aware model load and unload, surge handling, and a defensible pricing model. Lift-and-shift doesn’t cover it.
Key people
- Craig Schultz led the cloud-hosted initiative at RMS that I was architect for. He came from Lawrence Livermore National Lab, and became the hub of my 2011 to 2014 consulting cluster: NeoPrime, Zanttz/ShadowNetworks, QCRI, and a counter-WMD field exercise.
- Vladimir Syerik, an engineer Paul had hired at GoldMail, became a close friend. Both later applied to RMS. The hiring manager interviewed Vladimir first and asked about Paul, Vladimir gave high praise, and Paul got the role.
- John Compton, a linguist and analyst Paul worked with, later joined Paul on the Qatar/QCRI engagement.
Technologies used
C# and Java core services; hosted SaaS re-platforming of a desktop modeling binary; exceedance-probability catastrophe models over very large data models; a compute-scheduling and memory-management layer with prioritized multi-request scheduling, load-on-demand and unload across multiple models, and scaled shared infrastructure; a usage and pricing model for hosted heavy computation; surge and on-demand recompute; secure multi-tenant service delivery.
Patents / awards
None.