Uh-Huh: A Voice Agent for Busy-Handed Workers

A plumber under a sink, a mechanic elbow-deep in an engine, a shipper on a motorbike — each still has to answer the phone or lose the job. Vietnamese delivery riders reportedly take calls mid-ride as an accepted daily hazard, "đặc thù công việc" (tuoitre.vn, 2024-04-17). US trades businesses pay $250–$2,100/month for a human to answer every call so it never hits voicemail (Smith.ai, Ruby.com) — one plumbing company recovered $1.33M in a year, $200K in one month, once it stopped losing after-hours calls left unanswered 10–11 hours (ServiceTitan / Brothers Plumbing, 2025–2026). Uh-Huh is built for that exact moment: the customer calls, and the worker can't look at a screen or type, but can grunt one syllable. AssemblyAI's real-time streaming transcribes the call live; the agent turns the request into a short question in the worker's ear ("Sink job, Tuesday 3pm — good?"). The worker answers "uh-huh," "nope," or "later" — a one-syllable lexicon that catches instantly over engine noise. That syllable becomes a booking card, a text back to the customer in the worker's name, and a tick on a "money kept" counter — calls that would've gone to voicemail, now booked. The twist against every incumbent found — ServiceTitan, Goodcall, Rosie, Smith.ai, Canary — is that all of them put the AI on the phone with the customer. Uh-Huh never does. The AI talks only to the worker; the customer always gets a real human's yes, never a synthesized voice. Not a better bot voice, but no bot voice at all on the customer's side. Honest limits, up front: the streaming pipeline here is English-only for now — Vietnamese streaming support isn't stated in AssemblyAI's docs and is treated as MISSING, not assumed. The one-syllable classifier is a small hand-built lexicon, not a general speech model — it will miss anything outside it. The demo's customer caller is a pre-recorded clip, not a live caller, since no real trades business's phone line was wired in during the build window.