A daily brief for luxury client advisors. It reads five tables and produces one ranked list: which client to approach today, with which piece, on which channel, at what time, in what tone — and for some clients, a recommendation to say nothing at all.
▶ Live app · 📄 Today's brief · 📓 Build notebook
The static brief rebuilds itself every morning at 07:00 IST via GitHub Actions — the date in its header is today's.
| File | Grain | Rows |
|---|---|---|
clients.csv |
one client | 15 |
products.csv |
one piece | 25 |
opportunities.csv |
one outreach play | 8 |
interactions.csv |
one conversation | 52 |
purchases.csv |
one receipt line | 60 |
Gates run before scoring and cannot be outvoted. Consent, cooling-off periods, frequency caps and explicit refusals are absolute. A system that scores first and filters afterwards will eventually let a persuasive score talk it past something a client asked for.
What survives is scored on six dimensions — relevance, timing, channel fit, exclusivity fit, relationship strength and contact restraint.
Every recommendation carries its evidence, citing the rows it used.
15 clients reviewed, 12 approaches proposed, 3 deliberately left alone, 1 flagged for advisor approval where a personal occasion conflicts with a cooling-off rule.
python -m pytest -q # 13 tests, ~1s
The claim above — that gates cannot be outvoted by a score — is true on the day it is written and quietly stops being true three refactors later. These hold it in place.
The occasion exception is where the care is needed. A time-critical occasion may waive cooling-off; it may never waive consent, never the frequency cap, and never silently. Those are three separate promises, so they are three separate tests:
| Test | What breaks it |
|---|---|
| An occasion cannot rescue a client who withdrew consent | The consent check is dropped from occasion_exception |
| An occasion cannot waive the contact cap | The frequency check is dropped from it |
| A waived cooling-off is always flagged for approval | The waiver text stops saying APPROVAL REQUIRED |
| A client-level block is never overridden by a score | Anything but cooling-off becomes waivable |
| Every refusal states its reason | A held-back client is returned with no evidence |
| No draft is written for a client not being contacted | A message is prepared that could be sent by accident |
| Running twice gives the same brief | A draft stops being a template |
Each was checked by breaking it. Removing the consent line from
occasion_exception fails the first; removing the frequency line fails the
second; changing the waiver text fails the third. A test that cannot fail is
not evidence.
This is a co-pilot, not an agent. It automates the noticing, never the saying. Drafts are templates rather than generated text, so every recommendation is auditable — an LLM would sit in the drafting layer, under the same tone rules, with the advisor still approving every send.
Maison Aurelle is a fictional maison and every client, piece and interaction in this repository is synthetic, generated from a fixed seed.