Agent Shula

Shula is an AI phone agent who calls companies on your behalf. She dials the number, navigates the IVR menu, sits on hold for as long as it takes, and only rings your phone when a real human is on the line.

She has a name and a personality because she does what a great human assistant would do — persistently, patiently, and on your terms.

Because the relationship matters. "An AI phone bot" is a tool. Shula is a teammate who fights the IVR maze for you. The personality makes the experience feel less like dialing a robot and more like sending a competent ally on a mission.

Practically, it also makes the UI more readable: "Send Shula" is more memorable than "Initiate outbound call sequence."

Shula is part of a larger product called GreenEye Agency. Today she handles outbound calls. Eventually she'll be one of several specialized agents under the same umbrella — with a sister agent who screens incoming spam calls, another who summarizes voicemails, and so on.

For now, when people say "Shula," they mean the outbound-call hold-fighter you're using right now.

Shula uses Anthropic's Claude as her brain — specifically Claude Haiku 4.5 for the live in-call decisions (because it's fast and cheap) and the larger Claude models for offline analysis and learning.

Telephony runs on Twilio. Speech-to-text is built into the call gather pipeline. The "is this a human or a robot?" decision is a hybrid: a fast keyword detector handles obvious cases, and the LLM judges the ambiguous ones with full call context.

IVR stands for Interactive Voice Response — the automated menu you hear when you call any large company: "Press 1 for English, press 2 for billing..."

It's called a maze because companies design these systems to deflect you: deep menu trees, AI gatekeepers ("In a few words, tell me why you're calling"), endless hold music, "high call volume" rejections, and dead-end loops. Most callers give up before reaching a human. Shula doesn't.

Worse than you think. Americans collectively waste an estimated 900 million hours on hold every year. The average customer-service call is 13 minutes, most of it silence and music. Two thirds of callers hang up frustrated.

Companies know this. The mazes are partly designed to reduce headcount cost — if you give up before reaching a human, they don't have to staff one. Shula tilts that economics back in your favor.

You give her a phone number and a one-line description of what the call is about. Then:

1. She dials the number through Twilio.
2. She listens to the IVR and presses the right buttons.
3. She talks past AI gatekeepers if needed ("I have a billing issue, please connect me to an agent").
4. She waits on hold — for minutes, hours, however long it takes.
5. The moment a real human says "Hi, this is Sarah from..." she calls your phone and bridges you in.

Your phone stays free the entire time. You only get involved at the end, when there's a human ready to talk to you.

Two layers. First, a fast keyword detector scores the audio: phrases like "thanks for holding" or "my name is" boost the human score; "please hold," "your call is important," and music cues lower it.

For ambiguous cases (which is most of them), the audio gets sent to Claude with the last 20 transcript entries as context. Claude decides: human, AI agent, hold message, IVR menu, or wait. The dual-layer design is fast for the obvious cases and accurate for the tricky ones.

Yes, sometimes she gets it wrong. Every call you give feedback on (👍/👎) helps tune the prompts.

If you set a voicemail message before the call (or pick a template like "callback request"), Shula leaves it automatically and ends the call.

If you didn't set one, she ends the call and marks the outcome as "voicemail" in your history. You can then click "Retry VM" on that history row to try again with a message.

Shula stops the call and sends you a "Shula needs your help" notification at the top of the dashboard. You enter the missing piece (account number, ZIP, etc.) and she'll have it on the next call.

You can also pre-load info on each contact: open the phonebook, edit a contact, and add stored info like SSN, account number, ZIP, etc. Shula will use it automatically next time. For numeric info, she'll enter it on the keypad (DTMF) instead of speaking it.

Functionally indefinite. She held on the IRS line for 108 minutes recently, correctly identifying every recorded hold message along the way, and connected the user to a real agent in the end.

You can cancel from the dashboard at any time if you change your mind.

Open the dashboard. Pick a contact (or type a number), tell Shula what the call is about ("billing dispute", "appointment change"), and hit Send Shula. That's it.

You'll see her status update live as she navigates the IVR. When she finds a human, your phone rings — pick up and you're already connected.

Yes. While a call is in progress, the "Send Shula" button turns into a red Cancel Call button. Hit it and she hangs up immediately and politely.

Web dashboard — works on any modern browser (desktop best, mobile usable).

Android widget — a native app for sending Shula and watching status from your phone home screen.

The phone Shula calls when she finds a human can be any phone — mobile, landline, doesn't matter.

Today this is a personal project, no commercial pricing. Behind the scenes each call costs roughly $0.25 to $2.00 — mostly Twilio per-minute charges (toll-free is ~$0.014/min) plus a few cents in LLM inference.

That 108-minute IRS marathon cost about $1.95 total. Compare to two hours of your life.

All your stored info (name, address, phone, SSN, account numbers) lives in a SQLite database on a private server — no third parties, no analytics, no marketing. The server is protected by HTTPS and basic-auth login.

During a call, sensitive info (like an SSN) is sent to the IVR using DTMF tones (the keypad), exactly as a human would. It's not sent to any LLM.

Yes. Three feedback loops:

1. Per-call feedback — thumbs up / down on every call tunes future LLM prompts.
2. IVR learning — when Shula successfully navigates a company's menu tree, she remembers the path so the next call is faster.
3. Training facility (coming soon) — we'll run Shula against a fixed roster of companies on a schedule, analyze where she stumbles, and tune.

The product gets noticeably better with every iteration of feedback.