A private project and a public account of that project are different systems. I publish enough to make the work understandable and keep household data, raw recordings, and identifying details out of the site.
Publication boundary
What I do not publish
I do not publish real household manuals, raw private audio, conversations, credentials, tokens, account names, private network details, device identifiers, precise locations, routines, faces, voices, mail, or personal documents. Public examples must be fictional; real artifacts must be sanitized and captioned before publication.
Analytics
The public site measures visits
AI-Buzz has no accounts, forms, comments, or search. On an ordinary visit, Google Analytics loads if your browser can reach it. There is no consent switch. The site does not use advertising features or session replay, excludes browsers it identifies as automated, and lets me mark my own browser so it stays excluded before analytics starts.
Before Google Analytics loads, the site removes the query string and page fragment from the address. Google Analytics receives the page path and title and a shortened referrer. For an AI-Buzz referrer, it keeps only the path; for an outside referrer, only the origin. Google may also receive your IP address, visit time, browser, device, screen size, and approximate location, and it may set _ga cookies. Advertising storage, advertising user data, personalization, and Google Signals are disabled.
This code runs only on the public AI-Buzz pages. The separate private House AI application does not load it, and public-site analytics does not receive House AI audio or transcripts. You can block Google Analytics by disabling JavaScript, using a content blocker, or blocking its network requests.
Project data
House AI keeps the real manual out of Git
If real household facts are added for the trial, they stay in a local file that Git ignores. The House AI gateway does not write a conversation database. That does not prove that the phone, operating system, model server, speech service, or private-network service leaves no logs; the complete working path still needs a live retention check.
Content-free turn metrics stay in page memory by default. An explicit local opt-in can append an allowlisted JSONL timing record. The current record excludes audio, transcripts, answers, manual content, identities, hostnames, and network addresses, but retention is still part of the unfinished live audit.
What local means
The phone is a client
The intended Galaxy browser captures audio and sends it over a private Tailscale connection to the Windows PC. Speech recognition, Qwen inference, household lookup, and speech generation run on that PC. A private connection does not mean inference happens on the phone, and it does not prove zero retention across every component.
Build work
Coding agents are a separate cloud-assisted tool
During research and development, ChatGPT or Codex may receive code, public sources, logs, or text that I deliberately provide and sanitize. They do not receive live House AI audio, real household facts, credentials, addresses, private device data, or unrestricted access through the voice page. A project’s local runtime is not the same thing as the tools used to build it.
Infrastructure
Hosting still sees a request
Hosting and network services may receive your IP address, browser details, requested page, visit time, and security signals so they can deliver and protect the site. Links to source projects and documentation lead to sites with their own privacy practices.