Kaku Privacy Policy

• The short version
• What Kaku collects
• Why we collect it
• The AI companies that read your messages
• Voice
• Phone calls
• Text messages
• Email
• Payments
• Where your data lives
• What Kaku remembers about you
• How long we keep it
• See, export, correct, delete
• Cookies and browser storage
• Children
• Security
• The laws behind this policy
• Changes to this policy
• Contact

The short version

• Kaku is an AI. Her replies are generated by large language models run by other companies. What you type, say, or upload is sent to one of those companies to produce each reply, along with what Kaku remembers about you.
• She remembers you across conversations on purpose. Her memory of you is a folder of plain-text files on servers we control, and our team can read it.
• We do not sell your information, and we do not use your conversations to train a model of our own. Each night Kaku does read summaries of everyone’s recent conversations, together, to update her own written notes about coffee, our products, and how to help. See The AI companies that read your messages.
• Nothing is deleted automatically today. Ask, and we delete it by hand.
• Kaku is instructed to say she is an AI whenever you sincerely ask.

What Kaku collects

Your account

[CONFIRM when built: Kaku accounts were designed on September 16, 2026 and are not yet live. Until they are, Kaku has no accounts of her own; on the Quichapa store she recognizes you through your store login.] When you create a Kaku account we ask for your email address, a name (what Kaku should call you), and a password. Your password is stored as a one-way hash; we cannot read it. We also ask for a mobile number, on the web and in the app alike, because Kaku texts you a code to confirm it is you and reaches you about your account by text. A profile photo is optional; see Where your data lives for how photos are stored.

Your conversations

• Text. Everything you type to Kaku and everything she replies.
• Voice. If you use the microphone, the audio from your microphone and the transcript made from it. Her spoken replies are made from her reply text. See Voice.
• Images and documents. Photos, PDFs, and text files you send her, on tiers that allow uploads. The image or file is used to produce the reply and is not kept in your conversation history afterwards; the history keeps a note that an image or file was sent. [CONFIRM: no code path saves uploaded image or file bodies to disk — the reader found none, but verify on the live server, including temp files.]

Usage counts

For every message, we record when it happened, which model answered, roughly how much text it processed, your tier, and an identifier for you. This is how the monthly allowance is counted. It does not contain the message itself. If you are not signed in, the identifier is the cookie Kaku sets in your browser; if there is no cookie, it is your IP address.

Device and time signals

When you open or close Kaku, or react to one of her messages, the widget tells her your time zone, whether you are on a phone or a desktop (worked out from your browser’s user-agent string), when you last chatted, and the first 200 characters of the message you reacted to. Kaku uses these to greet you at the right time of day and pick up where you left off. They are folded into her prompt for that conversation and are not written to a file on their own.

Activity on sites that host her

Some sites that host Kaku also run her activity tracker. Quichapa’s store does. There, the tracker records page views, product and product-image views, the variant you pick, searches and which result you click, cart additions and removals, checkout steps, orders you complete, which account pages you open, your answers in the subscription quiz and changes to a subscription, clicks, how far you scroll, when you start or stop voice, when you open or close Kaku, and when you return, tied to the same cookie or account. Kaku reads this each night so she knows what you were looking at. kakucoffee.ai itself does not run the tracker. [CONFIRM: the kakucoffee.ai page loads only the widget, not kaku-tracker.js — confirm this stays true, and check price-index.html too.]

If you call, text, or email Kaku

Your phone number, the recording and transcript of a call, the text of a message, or the sender, subject, and body of an email. Details in the Phone calls, Text messages, and Email sections.

Technical data

Your IP address, browser type, and the usual web-server request logs, used to keep the service working and to spot abuse. Our server process logs also capture some of what passes through: the transcript of a voice memo, each phrase you say on a phone call, and the first line of a text you send us. [CONFIRM: how long nginx access logs and PM2 process logs are kept on the server (logrotate settings), or remove the three console.log lines — kaku-voice.js:392, telnyx-voice-bridge.js:253, routes/telnyx.js:128 — and drop the second sentence.]

Why we collect it

• To answer you, and to remember you so you do not have to repeat yourself.
• To count your monthly allowance and, if you subscribe, to bill you.
• To keep Kaku safe: her guardrails log the rule that fired and the conversation it fired in, so we can see what happened.
• To hand a conversation to a person when Kaku cannot help. When she does, the last ten messages are emailed to our team.
• To improve Kaku. Our team reads conversations to see where she goes wrong. Each night Kaku also reads summaries of everyone’s recent conversations, together, and updates her own written knowledge about coffee, our products, and how to help; that shared knowledge shapes every conversation she has. We do not use your conversations to train a model.
• To send you account messages: verification codes, receipts, and notices about your subscription.
• Marketing only if you opt in, and you can opt out at any time.

The AI companies that read your messages

Kaku does not run on a model we built. Each time you send her a message, we send it to a model provider along with the last twenty messages of your conversation and a “system prompt” that contains Kaku’s instructions and her memory of you: your identity file (name, email, phone number if she has it, customer ID, and for wholesale accounts your business name and type, account status, payment terms, and average monthly spend), your preferences, her most recent notes about you, your order summary, any notes our team has written about you, and the handoff note she last wrote herself. The provider returns the reply. If you sent an image, the image goes too; if you sent a document, up to about 12,000 characters of its text goes with your message.

Today, Kaku’s conversations run on xAI’s Grok models. [CONFIRM: the live KAKU_PROVIDER value on the server. The code’s default is Anthropic; the architecture notes say Grok is live. The policy must name whichever is true on the day it publishes.] Kaku is built so that we can switch the model provider. The providers she can be switched to are Anthropic (Claude), OpenAI (GPT), Google (Gemini), or a model running on hardware we own. If we switch, we will update this page.

Whichever model answers your chats, these companies also handle parts of the work today:

• xAI also runs Kaku’s overnight work. Each night we send it the previous day’s transcripts to write a daily summary of each person’s conversations, and at longer intervals the daily summaries to write weekly, monthly, and annual ones. Each night it also receives your preferences, Kaku’s notes about you, your identity file, and your recent site activity so she can notice things worth following up. If you are signed in, every message you send is also checked by a fast xAI model for whether it contains something Kaku should remember. And each night Kaku reads the past week’s daily and weekly summaries of everyone’s conversations together and asks xAI to propose additions to her own general knowledge — about coffee, our products, and how to help — which then shapes every conversation she has with anyone. She is instructed not to carry names or personal details into that shared knowledge, but the summaries she reads do contain your name. [CONFIRM: recommended code fix — strip the “Customer: name (id)” header line in collectRecentSummaries() (kaku-self-author.js) before the summaries reach the prompt; then this sentence becomes “the summaries she reads are labeled by customer number, not by name.” The no-PII rule is a prompt instruction today, not code.]
• OpenAI turns Kaku’s summaries, notes, and order history into numeric “embeddings” so she can search her own memory. Your raw transcripts are not sent for this; summaries and notes are. [CONFIRM: EMBEDDING_PROVIDER is openai on the live server; the code can also use Google or a local model.]
• Anthropic reads email sent to Kaku’s mailbox so she can understand it: the sender’s address and name, the subject, and the message body (see Email). If xAI is unavailable, Anthropic also performs the memory check described above, receiving your last twelve messages and your preferences file.
• Deepgram turns your voice into text. Every request we send it carries Deepgram’s opt-out flag, so Deepgram does not use your audio or transcripts to improve its models and keeps them only long enough to process the request.
• ElevenLabs receives the text of Kaku’s replies, which can include your name and things you told her, and turns it into her voice. We have opted out of data use in our ElevenLabs account, so ElevenLabs does not use what we send to improve its models. [CONFIRM: ElevenLabs publishes no retention period for standard plans; ask them or leave retention unstated.]
• Google is not in use today. If we switch to it we will update this page first.
• Telnyx keeps call recordings for one year unless we delete them sooner (see Phone calls). It is a telephone carrier and does not train AI models on calls or texts.
• Brevo (email delivery) and Migadu (mailbox hosting) process email as our service providers and do not use it to train AI models. [CONFIRM: Brevo’s and Migadu’s data processing terms — the reader confirmed the AI providers above from their published terms but did not read these two.]

If we ever want to use your conversations to train a model, we will tell you first, ask you, and take no for an answer. We will not make that change by quietly editing this page.

[CONFIRM when built: Apple Guideline 5.1.2(i) requires the app to get your explicit permission, in the app, before sharing your data with a third-party AI. The account design has no first-run screen yet. Build one that names the current model provider, says your messages and Kaku’s memory are sent to it to generate replies, links this page, and requires a tap to continue; log the acceptance with the terms-version record. Do the same on web signup. Fill in the App Privacy label to match this page.]

Voice

When you talk to Kaku with the microphone, the audio streams from your device to Deepgram, which turns it into text. That transcript is then handled as a normal chat message, marked as spoken. We also send Deepgram a short list of coffee words so it hears them correctly. Kaku’s spoken replies are made by ElevenLabs, which receives the text of her reply and returns audio.

Voice memos work the same way: the recording is sent to Deepgram for a transcript. A copy of the memo audio stays in your own browser’s storage so you can play it back; our servers do not save the audio file. The transcript of a memo, like any message, goes into your conversation and also appears in our server process logs (see How long we keep it).

We do not use your voice to identify you, and we do not build a voiceprint. What Deepgram and ElevenLabs may do with audio and text is described under Training.

Phone calls

Kaku answers the phone for Quichapa Coffee at 435-586-7634. [CONFIRM: that this is the Telnyx number Kaku answers, and whether kakucoffee.ai will have a number of its own.] Calls are carried by Telnyx. When Kaku answers a call live she says at the start that the call is recorded, and the whole conversation is recorded. The recording is stored at Telnyx for one year unless we delete it sooner. It is transcribed by Deepgram, and the full transcript is kept in our call log with your number and the time of the call. A short excerpt goes into Kaku’s memory of you, and our team is notified with the transcript. During a live call Kaku uses a fast xAI model to respond. If we do not already know your number, Kaku creates a memory file keyed to it, holding the number and what you said, even if you have never been a customer.

Voicemails are recorded after the greeting and transcribed the same way. [CONFIRM: the voicemail greeting is Kaku-authored; check that it tells callers the message is recorded and transcribed.]

Our call log keeps the most recent 1,000 calls. [CONFIRM: whether we ever delete recordings at Telnyx before the one-year default, and who owns doing so on a deletion request.]

Text messages

We text you for three reasons: verification codes; notices about your account or orders; and, only if you switch on “News & offers (text)” in your profile, occasional marketing texts, which you can switch off there or by replying STOP. When you give us a mobile number you agree to this:

By providing your phone number, you agree to receive account and order text messages from Quichapa Coffee. Message frequency varies. Message and data rates may apply. Reply HELP for help, STOP to opt out. Consent is not a condition of purchase. We do not share your number with third parties for marketing.

[CONFIRM with counsel: the consent paragraph above covers account and order texts only. The store has a marketing-text toggle (off by default) and an admin-only bulk text tool whose description says “promotions.” Marketing texts need their own express written consent under the TCPA (47 CFR 64.1200(f)(9)) and their own registered use case with the carriers. Either remove “promotions” from the tool descriptions (tools/telnyx-sms.js:265, tools/brevo-sms.js:141) and have Kaku’s guardrails refuse promotional bodies — then cut the third reason above — or build a separate marketing opt-in with its own consent wording and register it. Never reword the quoted paragraph.]

Texts are sent through Telnyx, from Quichapa Coffee: even when Kaku is the one writing, the text comes from Quichapa Coffee’s registered number. [CONFIRM: the storefront’s Kaku tools can also send through Brevo’s SMS API — name Brevo here if that path is used for customer texts.] Texts you send to us are read by Kaku, stored in our message log (the most recent 2,000 messages), and summarized into her memory of you. If we do not already know your number, Kaku creates a memory file keyed to it, holding the number and what you said. Automated texts are limited to codes, account or order notices, and the marketing texts described above; anything Kaku writes to you herself is reviewed by a person before it is sent. We keep a record of every text we send you (your number, the time, and the source).

If you reply STOP, Telnyx blocks further texts to your number from ours. [CONFIRM: no code on our side keeps an opt-out list — routes/telnyx.js treats every inbound text the same and tools/telnyx-sms.js never checks one. This sentence relies on Telnyx’s automatic STOP handling on the messaging profile for +1 435-586-7634; confirm it is enabled there. Better: build a suppression list checked in sendTelnyxRaw() so the promise does not depend on a carrier setting.]

The text-message program is registered with the carriers under Quichapa Coffee, and the consent wording above is the wording in that filing. [CONFIRM before kakucoffee.ai sends its first code: a text “from Kaku” to someone who has never dealt with Quichapa Coffee is a brand mismatch the carriers filter. Either (a) send every Kaku text from the Quichapa Coffee identity and show this exact consent paragraph at kakucoffee.ai signup, as this draft assumes, or (b) file a Kaku brand and campaign first. The filing was rejected twice before; do not reword the paragraph under any option.]

Email

We send email through Brevo: codes, receipts, subscription notices, and replies. When Kaku hands a conversation to a person, the last ten messages of that conversation are emailed to our team through Brevo, together with your name and email address. We keep a record of every email we send you (your address, the subject, the kind of message, and whether it was delivered) and a copy in our own mailbox’s Sent folder.

Email you send to Kaku is received in our mailbox (hosted by Migadu), read by Kaku, and interpreted with help from Anthropic, which receives the sender’s address and name, the subject, and the body. Kaku’s reading of it is saved to her memory of you. [CONFIRM: kakucoffee.ai does not have its own mailboxes yet; today this describes the Quichapa mailbox Kaku watches. Update when the kakucoffee.ai addresses exist.]

Marketing email only goes to people who opted in, and every one carries an unsubscribe link and our postal address. Anything Kaku writes to you by email is reviewed by a person before it is sent.

Payments

Subscriptions bought on the web are processed by Stripe. Your card details go straight to Stripe and never touch our servers. Stripe keeps your saved payment method so renewals can be charged; we keep your Stripe customer ID, your subscription status, and the dates of your billing period. Today the charge is made by Quichapa Coffee’s Stripe account. [CONFIRM: the statement descriptor customers will see on their card statement.]

When the Kaku app sells subscriptions inside the app, Apple handles the payment under Apple’s terms. We receive a token that identifies your purchase, not your payment details. [CONFIRM when built: In-App Purchase is planned, not live.]

Where your data lives

Kaku runs on servers we rent from DigitalOcean. [CONFIRM: that every Kaku server, including the one serving kakucoffee.ai, is a DigitalOcean machine, and which region.] Your memory files, transcripts, usage counts, and logs are on those servers.

Profile photos and wholesale company logos are stored on DigitalOcean’s file storage at a public address built from your customer ID. Anyone who has the link, or your customer ID, can view the image. Uploading a new photo replaces the old one at the same address; the old image is no longer served. We do not delete the image unless you ask. [CONFIRM: recommended code fix — use a random key or a signed URL in avatar-upload.js and remove the open /wholesale/avatar-url/:customerId lookup (kaku-agent.js:3511); then soften this to “at a private address only you and our team can reach.”]

What Kaku remembers about you

Kaku keeps one memory per person. It is a folder of plain-text files on our server, and it is what makes her feel like she knows you. Ours is not a black box; you can ask to read it. It holds:

• Identity: your name, email, phone number, and customer ID where you have one, and, for wholesale accounts, your business name and type, account status, payment terms, and average monthly spend.
• Preferences: brew method, roast, grind, favorite products, dietary notes, and anything you ask her to remember.
• Notes: Kaku’s running observations — things you mentioned, complaints, follow-ups she owes you. Both she and our team add to this file. For wholesale accounts, a nightly check adds a note when payment terms are overdue.
• Orders: your order history in plain English, how many orders you have placed, and how much you have spent, where she has access to a store you buy from.
• Transcripts: every message and reply, with timestamps, for people who are signed in. Visitors who are not signed in do not get a transcript file; their history lives only in their own browser.
• Summaries: daily, weekly, monthly, and annual summaries of your conversations, and a five-year portrait, written overnight. Each summary file is headed with your name.
• A handoff note Kaku writes to herself at the end of a conversation so the next one picks up where you left off.
• Activity: on sites that run her tracker, a daily log of what you looked at.
• Emails and texts you have sent her, as she understood them.
• A search index built from the summaries, notes, and orders so she can find things fast.

If you visit without signing in, Kaku still makes a small visitor file keyed to the cookie in your browser, so she recognizes you next time. When you later sign in, that visitor file is merged into your account and the old copy is archived, not deleted. If you call or text us, she makes a file keyed to your phone number in the same way, and merges it when she can match the number to you.

Your identity file, preferences, her recent notes, your order summary, team notes, and her latest handoff note go to the model provider with every reply. Transcripts, summaries, and activity logs are sent overnight for the summary and follow-up work described under The AI companies that read your messages. Our team can read all of it, to help you or to check her work.

How long we keep it

Honestly: today, nothing in Kaku’s memory expires on its own. Memory files, transcripts, summaries, activity logs, usage counts, guardrail logs, records of the emails and texts we sent you, our audit logs, and support tickets (kept in a single file on our server) are kept until you ask us to delete them. The call log keeps the most recent 1,000 calls and the text log the most recent 2,000 messages; older entries drop off. Telnyx keeps call recordings for one year. [CONFIRM: whether to set retention periods — for example raw transcripts and activity logs after a fixed number of months — and build the job that enforces them. This page can only promise what the code does.]

Our server process logs, which capture voice-memo transcripts, phrases spoken on phone calls, and the first line of texts you send, are kept [CONFIRM: the logrotate period for /var/log/coffeestore, or remove those console.log lines and delete this paragraph].

A live chat session is held in server memory for three hours and then dropped; the transcript file remains. The cookie Kaku sets lasts one year. What the widget stores in your browser stays until you clear it.

Billing records are kept as long as tax and accounting law requires, even after everything else is deleted.

See, export, correct, delete

You can ask to see everything Kaku has about you, get a copy, correct it, or have it deleted. Email office@quichapacoffee.com from the address on your account and tell us what you want. [CONFIRM: a kakucoffee.ai address for these requests once the mailboxes exist, so a Kaku customer never has to write to a roaster’s office address.] We will confirm it is you and reply within 30 days. We will not treat you differently for asking.

• See and export. We send you your memory folder as plain-text files, and your usage counts.
• Correct. Tell Kaku (“remember that I switched to a Chemex”) and she updates her notes, or email us and we fix it by hand.
• Delete. There is no delete button yet, so we do it by hand: your memory folder including archived merges, any phone-keyed file, and the search index; your transcripts and summaries; your activity logs; your usage counts; your entries in the call and text logs; the records of emails and texts we sent you and the audit-log lines that name you; your profile photo; and your account. We also ask Telnyx to delete recordings of your calls, and we remove your lines from our server process logs. [CONFIRM: the deletion checklist for each vendor — Telnyx recordings, and any retained inputs at xAI, Anthropic, OpenAI, Deepgram, ElevenLabs, Brevo — and who on the team owns it. This is a real workload; name the owner.] What survives: billing records we must keep for tax purposes, and, at Telnyx, the record that your number replied STOP so it is not texted again. Anything already sent to a model provider is subject to that provider’s own retention, described under Training.

[CONFIRM when built — launch blocker for the app: Apple Guideline 5.1.1(v) requires an in-app “delete my account” control before the app ships, and it must actually delete, not deactivate. The account design has it and a matching web control and a “what Kaku remembers” page to view, export, and forget; none is live. When they are, rewrite the Delete bullet above to describe the button and keep email as the fallback.]

Cookies and browser storage

Kaku sets one cookie, kakuVisitorId, which lasts a year. It is a random ID that tells Kaku it is the same browser, so she can recognize you between visits. It is sent with every message you send her.

The widget also keeps things in your browser’s local storage: your chat history if you are not signed in, your customer or visitor ID, your tier, a running count of credits used this month, when you last read or closed the chat, your profile photo address, and the audio of any voice memos you recorded. These stay on your device until you clear your browser data. Signed-in people get their history from our server instead.

When accounts launch, signing in will set a session cookie. [CONFIRM when built.]

kakucoffee.ai does not use third-party analytics, advertising pixels, or trackers. No third party collects information about your activity across other websites through Kaku. Sites that host Kaku have their own cookies and their own policies. [CONFIRM: no analytics script on kakucoffee.ai — the page loads only the widget today.]

We do not currently respond to browser “Do Not Track” or Global Privacy Control signals, because we do not track you across other sites or sell your data in the first place. [CONFIRM: nothing in the code reads these signals; California’s CalOPPA requires us to say so plainly.]

Children

Kaku is for people 13 and older. We do not knowingly collect information from anyone under 13. If we learn that we have, we delete it. If you are a parent and think your child has been talking to Kaku, email us and we will take care of it. [CONFIRM: the account design has an age gate; the FTC’s COPPA guidance says a neutral birth-year field, never an “I am over 13” checkbox. The Terms set 18+ to subscribe; see the age question there.]

Security

Connections to Kaku are encrypted. The keys we use with our providers live on the server, never in your browser or the app. Signed-in customers are identified by a token our server signs, not by anything typed into the chat. Kaku’s guardrails log anything that trips them. Access to the servers is limited to our team. [CONFIRM when built: password hashing, email verification, and the short-lived identity token for Kaku accounts are in the design, not yet live.]

No system is perfectly secure. If we discover a breach that affects your information, we will investigate promptly and tell you by email as soon as we reasonably can, as Utah law requires.

The laws behind this policy

We would rather tell you plainly which laws shape this page than pretend every privacy law in the country applies to a small roaster.

• Utah Consumer Privacy Act. It applies only to businesses with $25 million or more in annual revenue that also handle data on 100,000 or more Utah residents. We are far below that, so it does not apply to us today. We follow its structure anyway: what we collect, why, who we share with, and how to exercise your rights are all above.
• California. The CCPA/CPRA has revenue and volume thresholds we do not meet, and we do not sell or share personal information for advertising. California’s online privacy law (CalOPPA) does apply to any site that collects personal information from Californians, and this page is our notice under it, including the Do Not Track statement above. California residents can email us with any request about their data and we will honor it.
• Children (COPPA). See Children.
• Email (CAN-SPAM). Marketing email is opt-in, carries an unsubscribe link, and shows our postal address. [CONFIRM: the postal address to print in marketing email and in the Contact section below.]
• Texts (TCPA and carrier rules). See Text messages. Reply STOP at any time.
• Utah Artificial Intelligence Policy Act. Kaku never claims to be human, and she is instructed to say she is an AI whenever you sincerely ask. Her name and the widget make clear she is an assistant. [CONFIRM: this rests on a line in her core instructions (brains/kaku-core/08-honesty-and-epistemic-integrity.md:101), not on a hard guardrail — kaku-guardrails.js has no disclosure rule, and 01-identity.md calls her “human-esque,” which a model could echo. Build a guardrail (a pattern on “are you human / a person / a bot / an AI” that forces a disclosure line) and this can say “she tells you.” A visible “AI assistant” label in the widget header is the Act’s safe harbor and costs one line.]
• Utah breach-notification law. See Security.
• Apple. When the Kaku app is in the App Store, this page is the privacy policy linked from it, the app will ask your permission before your data goes to an AI provider, and it will offer in-app account deletion. [CONFIRM when built: both are App Review requirements (5.1.2(i) and 5.1.1(v)); neither exists yet.]
• Other states. As Kaku grows we will check each state’s privacy law and update this page. [CONFIRM as scale changes: several state laws have consumer-count thresholds; Texas has none for non-small businesses.]

Changes to this policy

When we change this page we update the date at the top. If a change matters — a new model provider, a new kind of data, a new use — we will email account holders before it takes effect. [CONFIRM: that account holders will get an email for material changes; the sending domain does not exist yet.] We will never change how your conversations may be used for AI training without asking you first.

Contact

Quichapa Coffee Inc.
Cedar City, Utah [CONFIRM: street or registered PO box address to print — CAN-SPAM requires it in every marketing email, and it should match here]
office@quichapacoffee.com
435-586-7634

[CONFIRM: a kakucoffee.ai contact address once its mailboxes are set up, so Kaku’s customers write to Kaku.]