https://docs.google.com/document/d/1VkowDoy3_s9TxPnSTHmZk6c9PrAD9MfwtbEYuRB71bs/edit?usp=sharing

Concierge

A voice-first travel planning agent built on Guava. Callers dial in and plan a trip out loud: Connie proposes hotels, restaurants, and experiences against a real budget, tracks the itinerary in SQLite, and offers concrete tradeoffs when something would push the trip over budget.

The distinguishing feature is that Connie matches how you talk. Rather than asking callers to pick a voice, she opens in a neutral register, listens to the caller's own words, and shifts to match them - a caller who opens with "yo, i wanna go to kyoto" gets a Gen Z register; one who opens with "Good morning, I would like to arrange a trip" gets a slower, careful one that repeats details back. Switches are silent, and an explicit request ("slow down") locks the style.

Project layout

concierge_app/
├── agent.py                  # guava.Agent(...) instantiation
├── main.py                   # entrypoint: --phone / --webrtc / --local / --chat / --roleplay
├── env.py                    # loads .env from the repo root into os.environ
├── logging_setup.py          # guava logging config
├── db.py                     # sqlite schema, seed data, traveler/trip/itinerary helpers
├── voice_styles.py           # the three registers + set_persona() composition
├── style_detect.py           # lexical register detection from caller speech
├── callbacks/
│   ├── lifecycle.py          # on_call_start (+ returning-caller recognition), on_session_end
│   ├── voice_style.py        # welcome task, on_caller_speech register matching
│   ├── questions.py          # on_question
│   ├── planning_actions.py   # on_action_request / on_action -> IntentRecognizer
│   └── profile_intake.py     # trip intake checklist + on_task_complete("trip_intake")
└── specialists/
    ├── hotels.py, restaurants.py, experiences.py   # inventory proposals
    └── budget.py             # category splits, tradeoff generation
concierge_dashboard.py         # judge-facing local dashboard (FastAPI), separate process
concierge_static/index.html    # dashboard frontend

Running it

Set GUAVA_API_KEY and GUAVA_AGENT_NUMBER in .env at the repo root (gitignored). --phone listens on whatever GUAVA_AGENT_NUMBER is set to.

guava run . -- --chat        # fastest iteration loop, text chat
guava run . -- --local       # real audio via laptop mic/speakers
guava run . -- --webrtc      # browser call via app.goguava.ai/debug-webrtc
guava run . -- --phone       # listen on GUAVA_AGENT_NUMBER - call it for real
guava run . -- --roleplay "You are a caller who wants a week in Kyoto"

Run the dashboard in a second terminal: uv run concierge_dashboard.py.

Voice styles

Three registers, all on the same grace TTS voice and the same identity (Connie / Concierge) - only the persona prose changes, via call.set_persona().

Style	Register
genz	Heavy slang, short sentences, reacts before recommending
friendly	Everyday American, warm, no slang - the default every call opens in
steady	Slow and clear, one option at a time, repeats names and numbers back

Detection lives in style_detect.py and is deliberately lexical rather than an LLM call, since it runs on every utterance: it has to be instant, free, and deterministic enough to unit test. It abstains rather than guessing when the evidence is thin, so a neutral caller simply stays in friendly.

To see recognition of a returning caller, call twice from the same number - the second call opens already knowing your name and your register.