Overview / Narrative Journeys
End to end · four stories
Narrative journeys — AI in action, end to end.
The stack is not a diagram. It is four threads of work that used to take days, or used to not happen at all. Each story follows a single input all the way to a real output — no hand-waving between the steps.
Four threads
Trigger, AI, outcome.
Each journey is the same shape: something happens, AI does the work, a real outcome lands. Hover any node to trace its thread.
Shoot → asset
Meeting → task
Question → answer
Idea → app
Journey 01 · Asset Ops
From a supplier shoot to a shoppable asset.
A supplier sends over a folder of photos and videos — 80 files, zero metadata, no naming convention. In any previous reality this lands with the marketing team as a manual cataloguing project. Today it lands in the pipeline.
Raw media lands in Drive
Photos and videos arrive in a shared Google Drive folder from the supplier — untagged, unordered, mixed quality.
Deduplication to real subjects
Asset Ops scans the folder and collapses variants to canonical subjects — the same product shot from three angles counts once, not three times.
Gemini writes the full record
Each image gets 14 structured fields: subject, tags, alt-text, suggested uses, colour palette. Videos get a timestamped shot list, reusable segments, and quotable moments.
Into the DAM with originals preserved
Originals go to Cloudflare R2, video goes to Cloudflare Stream, and typed metadata rows land in Directus — all linked, all searchable.
Marketing searches a live library
The team opens the DAM and finds every asset already described, tagged, and labelled with where it can be used — from day one.
Journey 02 · Hermes huxberrypm
From a meeting to tracked tasks.
A Google Meet ends. In a previous world, the action items live in a bullet-point document that nobody checks, assigned to nobody, due never. In this one, the meeting note is already being read.
Meet wraps with auto-notes
Gemini meeting summary is written automatically — decisions, action items, attendee ledgers. No manual note-taking required.
Hermes reads the notes hourly
The huxberrypm profile ingests new meeting summaries on a schedule — scanning for anything that looks like a decision or a commitment made to someone.
Task cards drafted from decisions
Each action item becomes a structured task card proposal — title, description, suggested owner, context extracted from the meeting thread.
One-tap review on Telegram
Proposals arrive in Telegram with inline approve/reject buttons. Nothing hits the board without a human tap. Control stays human; friction drops to seconds.
Approved cards on the Fizzy board
Approved tasks appear on the Fizzy project board, assigned and dated — ready to be picked up in the next standup.
Journey 03 · GBrain
Ask the ERP and CRM in plain English.
Someone needs to know the status of a deal, the current stock position, or what was decided in a call last week. In the old model they wait for a report, or ask the person who knows. In this one, they ask directly.
A plain-English question
"What are the open opportunities in hospitality right now?" — typed in chat, no special syntax, no knowledge of which system holds the answer.
GBrain answers from its index
Conversational questions — context, history, decisions, definitions — are answered from the 2,000+ pages GBrain has already indexed, with citations back to the source.
Exact numbers from the warehouse
For anything that needs a precise count or ranking, GBrain queries the Metabase analytics warehouse — sub-second SQL, no waiting on a scheduled report.
Freshest data straight from source
When the warehouse lacks a field or the question needs today's state, GBrain queries the live system directly — Odoo or NetSuite — and grounds the answer there.
A grounded, cited response
The answer comes back in plain English, citing which source it came from and at what freshness. Conversation continues — follow-up questions work too.
Journey 04 · Claude Code
From an idea to a deployed app.
A business need surfaces — track something, automate something, surface something for customers. In a company without in-house engineers, this used to mean a six-week agency engagement or an indefinite backlog item. Today it means a conversation.
A rough idea in chat
A plain-language description of what's needed — no spec, no wireframes, no ticket. Just the problem and the outcome that would solve it.
Claude Code reads the real systems
Before writing a line, Claude Code reads the codebase, the relevant APIs, and the GBrain wiki context — so the code it writes is grounded in how the systems actually work.
Working code, written and tested
Claude Code writes the implementation, iterates on errors, and verifies behaviour — running tests and inspecting output as it goes.
Previewed and confirmed in a browser
The running app is opened in a browser preview. Claude Code checks the visual output, clicks through flows, and confirms the behaviour matches the brief before shipping.
Live on a huxapps.com subdomain
Deployed to Cloudflare Pages or Vercel. A huxapps.com subdomain is pointed at it. The thing is live. This very site was built and deployed exactly this way.