A draft is only useful if it sounds like you — and only if it’s calibrated to who it’s going to.
Two questions every draft has to answer. How do I sound? — the four-axis voice. Who am I writing to? — the twelve relationship signals. Together they decide the shape of every single email Cadence writes for you.
Cadence reads how you write — not what you write about.
Each user uploads a minimum of five real, multi-turn email chains. The model breaks voice into four dimensions, encodes them in a per-user adapter, and applies them to every draft after.
Structural pattern
- Pleasantries first, or straight to the point?
- Short punchy paras, or detailed formatted lists?
- How do you transition between topics in one email?
Lexical footprint
- Industry jargon, sign-offs, colloquialisms
- Do you mirror the formality of your correspondent?
- Specific phrases unique to your voice
Relational register
- How do you express urgency without alarm?
- How do you disagree while staying warm?
- How do you signal something is non-negotiable?
Temporal rhythm
- How quickly do you respond, by thread type?
- Do you acknowledge before substantive reply?
- How do you re-open long gaps in a thread?
Hey Marcus,
Quick one — circling back on the rate card for the ZestyChicken Nachos brief. Need to lock this in by EOD Thursday so we can get contracts moving.
Totally get if usage is the sticking point — happy to talk through what flex looks like on our side. If you can share even a ballpark, I can take it to the client tomorrow morning.
Cheers,
Ava
The blueprint, made visible.
The same four dimensions, highlighted on a real approved email. Toggle a dimension to see what Cadence is actually learning from each draft you approve or edit.
- Greeting: shortened first name ("Marcus") · contact-specific
- Opener: "Quick one" — establishes time cost upfront
- Deadline framing: hard time, soft reason
- Negotiation tone: opens flex without committing to a number
- Sign-off: "Cheers" — AU-standard, warm, non-corporate
Twelve signals. Five clusters. One composite score that decides how every email gets written.
Cadence doesn't write to "the supplier." It writes to this person, with this history. After every reply you receive, twelve signals get re-scored. Each Sunday, an agent re-reads the last twenty threads and looks for the slow drifts the per-email updates miss. The score is the input to the four voice axes you already saw — structural, lexical, relational, temporal.
Do they execute reliably?
How cleanly do deals close?
The soft signals.
Long-term value.
Lifetime exposure.
Marcus at Lighthouse Media
First contact at Mantis Collective
Per-thread updates
Adjusts ≤±1 per exchange. Exceptional signals — broken commitment, unprompted referral — can move the score ±2.
Weekly drift agent
Every Sunday at 23:00 AEST, an agent re-reads the last 20 threads for every active contact. It looks for slow drifts — like a responsiveness score that slipped from 4 to 3 across five emails without a single one looking late.
From 5 emails to a continuously-improving private model — in 30 days.
The blueprint isn't a config file. It's a per-user adapter that's isolated, revocable and updated incrementally on every approved edit. Below is the flow, end to end.
Shared foundation-model fine-tuning
Train one big model on everyone's email and you get cross-contamination: Agency A's voice bleeds into Agency B's drafts; one supplier's NDA terms leak into another's quote.
- Cross-tenant contamination risk
- Cannot be revoked when a user offboards
- Updates require retraining the whole model
- Unacceptable for multi-tenant platform
LoRA per user — isolated, efficient, revocable
Each user's adaptation sits in its own cryptographic silo. The base model stays clean; only a small adapter swap is needed to draft in this user's voice.
- Computationally efficient — small set of weights
- Updated incrementally as new email data arrives
- Revocable entirely if a user offboards
- Stored database-per-tenant on AWS Sydney
It doesn't just read the email. It reads the deal.
Before a draft exists, Cadence has worked out what kind of email this is, who's writing it, and what commercial shape the deal is taking. 39 email types route across three bands — auto-send, queue, escalate — with the highest risk always winning. See the full 39-type taxonomy →
Who's writing — sender-role detection
Cadence reads the signature, domain and tone to tag the sender, then adapts register to who it's actually talking to.
What kind of brief — 7 sub-types
A brief isn't just a brief. Cadence distinguishes seven shapes so the reply matches the context it arrived in.
Every kind of creator deal, understood — and they stack.
From a one-off UGC clip to a whitelisted ambassadorship with event attendance and a spark code, Cadence knows the shape of the deal — because it's seen them all. One real-world deal is often several archetypes at once, across six families.
Brand-channel shoots, creator-channel UGC, collab and dual posts, co-produced shoots, studio capture, voiceover and photoshoot-only deals.
Whitelisting, TikTok Spark Codes, dark posts, and boosting an organic post into paid.
Affiliate revenue-share, CPA performance, hybrid base-plus-performance, ambassadorship retainers, equity partnerships and co-branded product.
Podcast integrations, live-shopping hosting, AR filters, newsletter mentions, long-form YouTube, tutorials, story takeovers and recipe/format development.
Event attendance, attendance plus content, hosting or MC duties, and trip or launch coverage.
Seeding with or without an obligation to post, exclusivity buyouts, licensing existing content, OOH/TVC likeness and strategic consultation.
It sounds more like you every week
Every edit you make teaches Cadence your greetings, sign-offs, tone markers and length by email type. Day 1, every draft needs a review. Day 30, it writes the way you'd have written it — and approval becomes a three-second glance.
It never loses a number
Every rate, counter and term is remembered with a timestamp and a source — across inbound emails and manual hub edits, in one unified history. See the Fact Ledger →
The AI is not always the right tool. Three conditions stop it entirely.
Confidence isn't the only signal. The platform also responds to explicit human takeover — for any thread, at any time. The principle: the AI works for the human, not the other way around.
All automation suspended for that thread until explicitly re-enabled.
Next action guided by the manager's verbal context — verbal record beats inferred state.
AI flags. No draft generated. Manager contacts the supplier directly.