Raghav Kattel

The request reaches a load balancer. Instead of one server doing everything, work is routed across several identical service instances, so no single process becomes the bottleneck — or the single point of failure.

One instance stops answering. Health checks mark it unavailable, the router stops sending it work, and the in-flight call is retried on a healthy instance while a replacement is scheduled. Nothing magical — redundancy, health checks and retries, designed in from the start.

When a request needs judgement, it moves through an AI pipeline: the input is embedded, relevant knowledge is retrieved, a context window is assembled, and a model streams back a response. The engineering lives in retrieval, grounding and evaluation — not in the magic.

Some requests must be verifiable by parties who don’t share a server. The request is signed and becomes a transaction: submitted, validated, included in a block and confirmed. Account abstraction lets an application sponsor and batch those steps, so the person on the other end never has to manage gas or keys.

Infrastructure is only interesting when it reaches people. At Rumsan I worked on Rahat, an open-source platform for humanitarian cash and voucher assistance, where a confirmed transaction meant support could reach families faster.