Pulse wasn't designed on a whiteboard. It was extracted from two real events that had to work the first time, then hardened into a platform on a third.
The engagement that forced edge inference into existence. White Night is a one-night public arts festival with no chance to patch quietly on night two, calling for an interactive AI experience at a venue with no reliable internet — ruling out any architecture that needed a live round-trip to cloud inference. Built in 3 weeks, bespoke and one-off; a load test against 500 simulated concurrent users surfaced a latency spike from a synchronous request path stalling under concurrency, so the queueing model was redesigned overnight to asynchronous batch processing to degrade gracefully instead.
White Night's result led to a referral from the Singapore Tourism Board, this time against a different failure mode: the compute economics of generating personalized content live for every visitor, which didn't pencil out at tourism scale. The fix was a hybrid system — pre-compute ~80% of content ahead of time by visitor segment, reserve live generation for the unpredictable 20% — the split Pulse is still built around today, validated with a blind A/B test to confirm visitors couldn't tell the two apart.
A separate engagement, same latency discipline: profile before you optimize. The rendering pipeline was profiled under 200 simulated concurrent users, surfacing three fixes — serial asset loading, no priority ordering between current vs. next content, and shared resources rebuilt per-request instead of reused. Applying parallel loading, a priority queue, and resource reuse cut end-to-end rendering latency from 2.3s to 380ms, a ~6x improvement visitors could feel standing at a booth.
Three field engagements in just over a year converged on the same patterns — edge inference, hybrid pre-generation, profiling-driven latency fixes. Rebuilding those from scratch per client wasn't sustainable, so starting in 2023 EFFOMA formalized them into Pulse, a reusable AWS-native platform. The hybrid pre-generation approach proved the most portable piece, since reused as a template across 3 subsequent EFFOMA projects, each with its own visitor-segment model but the same pre-compute/live split.