A week-1 stack audit

Foundation first.
Test every AI ask against it.
Sequence the rest.

Paste a client's discovery notes. It pulls the findings, gates each Gen-AI ask against the data foundation it needs, and hands back a phased roadmap a CMO can argue over.

Live engine

Pick a case, edit the notes, run the audit.

Client discovery notes

Client: Velora, regional beauty + personal-care. Six SEA markets: Singapore,
Malaysia, Indonesia, Philippines, Thailand, Vietnam. Regional creative team in
SG produces the master campaign assets; country teams/local agencies adapt.

Discovery notes:
- Salesforce Marketing Cloud is the "CRM/CDP" on the slide, but in practice
  only SG and MY are using the customer-data side of it. ID/PH/TH/VN are still
  working out of spreadsheets and local-agency exports. No shared hygiene
  process that anyone could point to.
- Country teams do not really trust the Singapore-built audience segments.
  One market lead said the segments "look clean in the deck, not in-market",
  so they usually rebuild their own lists before launch.
- Master creative is made in Adobe Creative Cloud. Finished assets, cutdowns,
  pack shots and translations live in a shared Google Drive everyone calls the
  DAM. No tagging, no real versioning. People search by filename or ask the
  designer on Teams. Old files occasionally get picked up by mistake.
- Social scheduling is split: Sprinklr for SG/MY, Later for ID/PH, and TH/VN
  use a local tool nobody outside that cluster has logins for. Need to confirm
  whether it exports cleanly.
- Paid media is Meta, TikTok and Google across markets. Shopee/Lazada media is
  run by ecommerce and comes back separately, usually as another sheet.
- Reporting pack is painful. Every market reports campaign results in its own
  format, then regional spends days normalising the numbers.

CMO / regional asks:
1. "AI that personalises our creative for each market" this year. Prompted by
   a competitor example; she wants to show progress before the next planning
   cycle.
2. Content production is too slow. Localisation and re-cutting take 2-3 weeks
   per campaign because every market starts from the master assets and makes
   variants by hand. CMO wants Gen-AI to speed this up.
3. Someone on the regional team floated an AI chatbot for customer product
   questions. Feels less owned; current knowledge base seems to be product
   pages, FAQs and whatever the care team has locally.

Pick an example or paste your own notes, then Run audit for a live extraction by gpt-5.4.

Method

What the model does, what the look-up adds, what the engine does

Four stages, left to right. The look-up is its own step between the model and the engine — which also fills the full width.

Model · merit

Reads the messy intake and pulls out findings — layer, type, dependencies, and a merit call on each AI idea. It does not sequence.

Evidence · prior work

Looks each finding up against prior engagements. Confirms a declared dependency, or surfaces one the brief never mentioned.

Engine · readiness

Builds the blocker map, gates every Gen-AI idea on its dependencies, and overrides the model where the foundation isn't ready.

Output · plan

Three phases. Foundation first, ready quick wins in parallel, gated ideas after their blockers clear. Deterministic from the same findings.

Merit and readiness stay separate. The model judges merit; the look-up grounds it in prior work; the engine judges readiness and sequences the plan. That logic lives in engine.py, the records in corpus.py, both covered by evals.

Framework

Four layers. One gate.

Gen-AI capability

generation, variants, optimisation, assistants

Activation channels

media, CMS, social, email, e-commerce

Integration & handoffs

do the tools talk to each other?

foundation

Data foundation

customer, content, market data

foundation

Every Gen-AI ask is gated

Ship now. Layers 1–2 are sound. Pilot.
Fix foundation first. Real opportunity, blocked by the layer beneath it.
Re-scope. A useful AI play exists nearby; the named one isn't it.
Don't pursue. Wrong-shaped, or value too ambiguous to defend.

The roadmap follows the blocker map, not the wishlist.

Future State

How this runs in production

The live demo implements the three highlighted steps; the rest is the automated pipeline around them — same boxed language, LIVE badges, a diamond for the human gate.

Future State

Discovery calls CCaaS · recordings → transcription

Stack inventory connectors · CDP · DAM · CMS · media · commerce

Canonical record one engagement record · backend

PII / client-confidential redaction privacy by design — before any model call

LIVE LLM extraction messy notes → typed findings

LIVE Prior-engagement retrieval RAG in prod · fixture-backed here

Prior-engagements library the look-up layer

LIVE Readiness engine gate · sequence · override — deterministic core

LIVE Phased roadmap foundation first · 3 phases

Agentification specs scaling blueprint · drafted by the model

Human review HITL · architect + client approve

Into existing rituals Slides · Notion / Confluence · Jira / Linear backlog

approved

shipped work feeds the library

Live in this demo Production pipeline around it

Discovery calls are transcribed and the stack is pulled in over connectors; everything lands in one engagement record, PII is redacted before any model call. Then the three live steps run — the model extracts findings, the look-up grounds them against prior engagements, and the deterministic engine gates and sequences the plan. That fans out into the phased roadmap and an agentification blueprint, a person approves before anything ships, and the shipped engagement feeds back into the library — so the look-up gets smarter each time.

Foundation first. Test every AI ask against it.Sequence the rest.

Live engine

What the model does, what the look-up adds, what the engine does

Four layers. One gate.

How this runs in production

Foundation first.
Test every AI ask against it.
Sequence the rest.