Hesper

A voice-first AI caddy that answers “what club?” out loud, in a caddy’s own voice.

Overview

Golf apps either drown players in dashboards or hand over distances with no judgment. Hesper is a native iOS app the player talks to: say “Hey Hesper,” ask about the shot ahead, and hear back a reasoned club recommendation in a chosen caddy personality, voiced through AirPods.

What I Built

• Designed the product end to end (brand identity, three caddy personalities, the full onboarding-to-round flow) and directed the engineering build through AI-assisted development.
• Specified a deterministic rules engine, in pure Swift, that computes yardage, wind, elevation, and lie adjustments and picks a club, so the math never depends on the LLM.
• Layered an LLM persona (Claude) on top of the rules engine to narrate recommendations conversationally and field follow-up questions, with ElevenLabs text-to-speech and Apple Speech powering the voice loop.
• Architected a five-stage voice pipeline (wake-word, speech-to-text, reasoning, text-to-speech, audio) targeting under 1.5 seconds end to end, with an on-device fallback for dead zones.
• Built a simulator mode for indoor golf bays: the player narrates the shot and Hesper computes plays-like distance and club with no GPS or course data required.
• Shipped a 45-plus-test unit suite covering the rules engine, course and hole detection, and voice-pipeline logic.

Tech

Swift, SwiftUI, SwiftData, Claude API, ElevenLabs, Apple Speech framework, Picovoice Porcupine (wake word).

Status

Working prototype; the voice loop runs end to end on the founder’s iPhone. Roadmap targets a TestFlight beta in golf-simulator bays before an on-course, course-data-dependent v1.0.