Human Perspective.
Platform Blueprint.
The working Azbro beta combines reviewed public knowledge, Contentstack-native conversation and an editable website. This blueprint distinguishes tested capabilities, current limits and the work still ahead; it does not claim a fully autonomous end-to-end platform.
From Human Knowledge to Human Connection
- 01 / Sources Identified
Registered Sources
Publications, podcasts and approved uploads
- 02 / Weekly Substack and Personal LinkedIn Checks Live
Draft Import
Agent OS fetching, attribution and change detection
- 03 / 41 Approved Source Records Published
Owner Review
Approve the public version; preserve speaker context
- 04 / Approved Sources, Site Copy and Episodes Live
Published Content
Headless CMS delivers approved page content
- 05 / Native Chat and Citations Tested
Cited Conversation
Approved evidence → Agent OS → Brand Kit voice → Linked citations
- 06 / Browsing Live
Meaningful Connection
Explore the work and choose the next step
The Promises Shaping the Build
- Zero required ongoing spend, within finite free allowances.
- Contentstack first, with a useful role for each included capability.
- Approved sources, explicit attribution and honest uncertainty.
- Minimal collection, explicit choices and optional browser conversation history.
- A useful website even when AI is unavailable.
One Platform. Four Connected Stories.
Content is the foundation. Automation moves it, identity gives it context, and delivery turns it into an experience. These reference views show how the pieces connect, with working and planned boundaries made explicit.
Model the Knowledge. Shape the Experience.
Page entries hold the words visitors see; shared settings hold reusable interface copy. Public Source entries keep the evidence distinct from presentation. Approval and publication are separate gates.
Page Content
Homepage · Architecture · Explore the Work
Shared Settings
Navigation · Help · Music · Analytics
Public Sources
Canonical URL · Attribution · Approved Text
Published Delivery
Read-only API → Next.js → Visitor
Working: three editable CMS pages, shared settings, CMS artwork, five episode links, restricted draft preview and 41 approved published source records. The source library and conversation use approved published evidence; original CV/profile PDFs and private contact details stay out of the public site.
From New Material to Reviewed Knowledge.
A scheduled workflow discovers registered publications and prepares material for an owner’s review. A successful fetch is not approval, and imported words are never automatically presented as Az’s own.
Registered Collections
Substack · Podcast · Website · LinkedIn
Scheduled Discovery
HTTP Fetch → Parse → Canonical URL
Draft Preparation
Duplicate Check → Attribution → CMS Draft
Owner Decision
Review → Approve → Publish
Live native workflows: chat gateway, weekly Substack intake and weekly personal LinkedIn checks. Native action tests created an unapproved feed revision and a newly discovered LinkedIn article. Canonical URLs and fingerprints distinguish new items, changes and existing drafts. The LinkedIn index is partial and each run checks at most three full articles. Company and website monitoring remain incomplete; scheduled whole-flow execution still needs observation. Nothing automatically approves or publishes.
A Digital Shadow With a Human Centre.
Identity and cultural context guide the voice: Yorkshire warmth, ambition, a dry sense of humour and respect for the people behind the work. Brand guidance shapes generation; the website’s visual design is implemented in code.
Identity and Culture
Professional Context · Values · Perspective
Yorkshire Voice
Plain-speaking · Ambitious · Dry Wit
Knowledge Vault
Identity Reference · Whole-Vault Retrieval Off for Public Chat
Grounded Response
Voice Guidance + Approved Evidence → Cited Answer
Tested: managed read-only Brand Kit connection, Yorkshire Voice and native generation. Public chat supplies approved CMS evidence with knowledge_vault disabled. Valid citations help inspection but do not certify factual accuracy.
From a Git Commit to a Public Beta.
GitHub holds the reviewed Next.js application. Launch builds and hosts it; Cloudflare connects the public hostname. Published CMS content is read at runtime so editorial changes do not require a code deployment.
GitHub
Versioned Code → Tested Branch
Contentstack Launch
Build → Deploy → HTTPS Hosting
Cloudflare DNS
azbro.zaver.co.uk → Launch CNAME
Visitor Experience
Published Content · Cited Chat · Optional Consented GA4
Live: custom hostname, HTTPS and GitHub-connected deployment. Preview credentials stay server-side; management credentials do not enter the website deployment.
One evidence model, several ways in. Each source needs its own discovery method, attribution and refresh policy. These flows show the current beta and the next connections without treating a working import as a finished agent.
Public LinkedIn Profile
- Public Profile
- Owner-Verified Snapshot
- CMS Review
- Approved Profile Context
The owner-supplied LinkedIn PDF has been extracted into two reviewed public records: a dated career profile and an About, skills and education addendum. LinkedIn chronology governs role dates and current roles where the CV differs. Original PDFs, private contacts and uncertain employment relationships are excluded. Future changes require a refreshed snapshot and owner review.
Open the Source ↗ (opens in a new tab)LinkedIn Articles by Az
- Weekly Public Profile Check
- Bounded Article Validation
- New or Changed Draft
- Owner Review and Publication
Sixteen personal articles are owner-approved and published. A separate weekly Sunday 09:00 UTC workflow discovers attributed links from the public profile, prioritises unseen URLs and rotates checks of known material. At most three full articles are fetched per run with redirect, author, canonical URL and date checks; fingerprints prevent repeated drafts. A native test discovered the new Azbro article, created an unapproved CMS draft and, after explicit owner review, that article joined the published library. The index exposed ten article links during testing, so this is partial discovery rather than proof of archive completeness. Comments and related articles are excluded.
Open the Source ↗ (opens in a new tab)MACHINE ACTOR. Website
- Public Website Index
- Resolve Original Publication
- Canonical Deduplication
- Review Source Context
The website helps discover the original episode, article or channel. Its links do not create a second copy of material already found on Substack or Spotify. A recurring index check and contextual source registry remain to be connected.
Open the Source ↗ (opens in a new tab)Substack Writing and Episode Guides
- Weekly RSS Check
- Parse Writing and Guides
- Compare Fingerprints
- Unapproved CMS Draft
Five episode posts, five guides and one Field Notes post are approved and published. The weekly Sunday UTC workflow checks the registered feed for new or changed material. Native parser and duplicate-filter tests passed with contextual excerpts and five Spotify-linked episode candidates; its CMS create action produced an unapproved revision linked to the existing record. Future changes remain drafts for owner review. Feed coverage does not establish complete archive coverage, and the first scheduled whole-flow run after this upgrade remains to be observed.
Open the Source ↗ (opens in a new tab)Podcast Episodes and Spotify Console
- Podcast Feed Discovery
- Match Spotify Episode
- Review Notes and Speakers
- Publish Console and Evidence
Five episodes have CMS-managed dates, descriptions and outbound Spotify links. Five episode posts and five guides are approved published evidence, alongside five owner-approved automatic-caption chunks covering Episodes 1–4. Captions are labelled potentially imperfect and have not been speaker-verified. Episode 5 captions remain unavailable. The console can merge published Spotify-enriched source records with the curated catalogue; the upgraded scheduled enrichment path still needs an end-to-end run.
Open the Source ↗ (opens in a new tab)MACHINE ACTOR. LinkedIn Articles
- Registered Company Listing
- Accessible Article or Export
- Preserve Author Context
- CMS Draft and Review
Six company articles have verified full text, publication dates and attribution, and all six are owner-approved and published. Two of the eight initially expected items remain unidentified. LinkedIn and Substack versions retain separate provenance. Scheduled company discovery remains incomplete; company publication does not establish that every statement was authored by Az.
Open the Source ↗ (opens in a new tab)CVs and Additional Owner Material
- Owner Upload
- Extract and Classify
- Review Public-Safe Text
- Approve and Publish
The owner-supplied CV and LinkedIn PDF produced three approved public records: career and experience, the LinkedIn career profile, and an About, skills and education addendum. The owner selected LinkedIn’s chronology and current roles. Public-safe text preserves provenance; original PDFs and private contacts remain excluded. Future uploads follow extraction, reconciliation, owner review and publication.
Import Once. Review Deliberately. Publish With Context. Canonical URLs prevent duplicate imports; fingerprints detect changes. Forty-one records have owner approval and publication. Only approved published evidence is eligible for public answers. New imports and changed-source revisions remain unapproved drafts until reviewed.
What’s Real. What Comes Next.
The beta website, with accessible routes, local fonts, supplied artwork and a server-only CMS reader.
The Homepage, Architecture and Explore the Work page models supply editable content. Site Settings manages shared navigation, footer, help, music and analytics wording. Restricted Visual Editor previews keep drafts private; the public site reads published content only. Page structure and visual styling remain in reviewed Next.js code.
GitHub-connected Next.js deployment on Contentstack Launch. The Beta is publicly accessible at azbro.zaver.co.uk. Cloudflare DNS points directly to Launch; HTTPS and the custom hostname have been verified.
Agent OS runs the managed native-chat gateway and separate weekly source checks. The Substack parser produces contextual excerpts and Spotify episode links; native step tests created an unapproved revision. Personal LinkedIn discovery runs Sundays at 09:00 UTC, validates at most three full articles per run and created a draft of a newly discovered article in its native test. The public profile index is partial, not a complete archive. Company LinkedIn and website monitoring remain incomplete. Imports never auto-publish.
The managed Agent OS connection uses read-only Brand Kit access and the Yorkshire Voice profile. Approved source passages are supplied explicitly; whole Knowledge Vault retrieval is disabled for public chat. Brand guidance shapes voice, while the website styling remains code.
The Next.js endpoint retrieves approved published evidence and calls Contentstack-native generation through Agent OS. A frontend response with linked citations has been tested. Citation IDs are checked against retrieved sources; this does not establish claim-level accuracy. Rate controls are per application instance, not a shared global quota. Remaining AI credits, per-answer credit cost and sustainable visitor capacity are not verified. The beta permits 20 requests per address per ten minutes, two concurrent requests and 100 requests per hour per application instance. Rate-limit responses explain when to retry.
Google Analytics loads only after acceptance. Fixed page-view and approved outbound categories exclude chat text. Reject or withdraw in Analytics Settings; consent choices expire after 180 days. Property receipt and duplicate-event verification remain owner-deferred.
A CMS review extension for attribution questions, duplicate sources and batch approval.
How Knowledge Becomes an Answer
Registered Source → Draft Import → Owner Review → Approved Publication → Retrieval → Cited Answer
Forty-one owner-approved source records are published: eleven Substack publications, sixteen personal LinkedIn articles, six company LinkedIn guides, five automatic transcript chunks covering Episodes 1–4, and three career/profile records. The newest-first library offers content-type filters and contextual excerpts. Chat retrieves approved published evidence, keeps dated career context together for follow-up questions, and resolves returned citation IDs to known source links. Automatic captions may be inaccurate and do not establish speaker identity; guest statements are not treated as Az’s own.
Hosting the Beta
Launch builds and hosts the Next.js application from GitHub. Published CMS content is read at runtime. The server keeps scoped delivery and preview credentials, plus the authenticated chat-gateway configuration, out of public browser bundles; CMS management credentials are not deployed to the website.
The public address is azbro.zaver.co.uk. Its Cloudflare CNAME uses DNS Only routing to Launch, and the website has been verified over HTTPS. The Beta badge makes the site’s stage explicit.
A Voice With a Human Behind It
Azbro’s Brand Kit captures plain-speaking Yorkshire warmth, ambition, dry wit and occasional dark humour. Its work principles centre on having a laugh, meaningful connections and mutual respect; ambitious execution and technical excellence that raise the floor and the ceiling; and results that help people and communities and earn their trust.
The Voice Profile uses Casual formality, an Informative tone, Subtle humour and Straightforward language. The managed Agent OS read-only Brand Kit connection has been tested. Public chat supplies reviewed CMS passages explicitly and leaves whole Knowledge Vault retrieval disabled. CSS implements the palette and typography; Brand Kit does not automatically theme the frontend.
Native generation and a frontend answer with linked citations have been tested. Answers are guided towards direct facts, useful context and restrained humour. Grounding remains an ongoing quality check: a valid citation ID does not prove every sentence is supported. Unknowns, guest attribution and historical context must remain explicit.
The Conversation Beta
The industrial messenger opens as a full-page conversation with clear Close and Explore controls, a prominent hero entry point, avatars, source-specific questions, cancellation and linked citations. Reply guidance adapts to factual questions, deeper explanations and collaboration enquiries; follow-up questions are optional. Loading animations represent a real pending request and respect reduced motion. Visitors can copy Az’s public email address into any email service or open his LinkedIn profile. Contact choices never replace the chat composer or require an installed email app; nothing is automatically sent.
Messages and recent context are sent through a server-side endpoint to Contentstack Agent OS and native Brand Kit generation. Processing may appear in Agent OS execution logs. Optional browser history requires opt-in, expires after seven days and keeps at most 30 messages; Forget removes the browser copy, not platform execution logs. There is no automatic paid-model fallback.
A CMS-Controlled Soundtrack
The CMS soundtrack kill switch is currently off, so the public page does not load a YouTube player. If enabled later, the visible player loads only after a visitor chooses it, and Stop unloads it. Playback availability depends on YouTube and the selected recording.
Built for People and the Agents Helping Them
Person and WebSite structured data, canonical/social metadata, robots.txt and a sitemap describe the public website. llms.txt, the Markdown profile and the public JSON profile use approved published information; their conversation status follows the enabled CMS interface and server configuration.
Supporting browsers register three read-only WebMCP tools: get_az_public_profile, list_az_approved_sources and get_azbro_capabilities. They expose public information and availability, with no messaging, consent, playback or actions on Az’s behalf. Tool contracts and cancellation are tested; native browser-agent interoperability remains unverified. This is browser integration, not a public remote MCP server.
Your Analytics Choice
Google Analytics is optional and loads only after acceptance. Rejecting leaves it off; Analytics Settings lets you withdraw. The browser stores a versioned choice for up to 180 days, then asks again. Old unbounded choices require fresh consent.
After acceptance, GA4 receives page views and approved outbound-action categories such as professional profile or podcast site. Page URLs exclude query strings and fragments, and we do not send chat text, names or email addresses. Google may process technical information such as browser and network data when the tag loads. Advertising consent and Google Signals are disabled. Analytics cookies are configured for 90 days.
Chat messages and recent context are processed by Contentstack Agent OS and may appear in execution logs. Optional browser conversation history is separate from analytics: opt in to keep up to 30 messages for seven days, or forget the local copy. The Privacy page explains these choices. Following an external link takes you to that publication’s own website.