Building Sonic Santa

Talking to Santa is a small piece of childhood magic that technology has mostly ruined rather than improved. Pre-recorded calls are generic, video services are scripted, and anything genuinely personal usually means a parent quietly doing an impression down the phone.

What was missing was a call that felt real: a conversation, not a recording, where Santa responds to the child in front of him rather than to a script written for everyone.

SonicSanta lets a parent set up a call in a couple of minutes. They add the child's name, a few interests, and one or two achievements from the year — a swimming badge, a new baby brother, a hard term at school.

When the phone rings, Santa already knows. The conversation is short, warm and specific, and it ends before the magic wears thin.

A parent-facing web application collects the call details and schedules the call. At call time a telephony layer bridges the child to a real-time AI voice pipeline: speech in, model reasoning in the middle, synthesised Santa voice out.

The personalisation lives in the prompt context assembled just before the call, so nothing about the child is baked into a model or a permanent transcript store.

The stack is deliberately boring where it can be, and interesting only where it must be. A static front end sits behind a CDN, scheduling and call orchestration run as serverless functions, and the voice pipeline is the one genuinely latency-sensitive component.

Round-trip latency is the whole product. A Santa who pauses for three seconds is not Santa. Most of the engineering effort went into streaming audio in both directions and overlapping the speech-to-text, reasoning and text-to-speech stages rather than running them in sequence.

Children's data raises the bar on everything. The design principle was to hold as little as possible for as short a time as possible: personalisation details are supplied by the parent, used for one call, and not retained beyond what is needed to deliver and support that call.

The conversational model is tightly constrained. Santa stays in character, stays on topic, does not ask questions that gather new personal information, and hands anything unexpected back to a safe, scripted path.

SonicSanta has an unusual load profile: near-zero demand for eleven months, then an enormous, concentrated peak in the fortnight before Christmas — largely between 5pm and 8pm.

That shape rules out fixed capacity. Everything is built to scale from nothing and back down again, with call scheduling used as a natural smoothing mechanism so peak evenings are spread rather than stampeded.

Perceived latency matters more than model quality. A slightly simpler response delivered instantly beats a better one delivered late.

Constraining an AI voice application is more work than enabling it. The interesting engineering is in the guardrails, the fallbacks and the failure modes — not the demo.