Platform plan

The new MMW platform

What we are building, why, and how it ships by the end of 2026.

Presented by
Matthew Blode
Date
16 November 2026

What MMW does

How a page gets marked.

Students write on paper. AI drafts the marking against the teacher's rubric. A teacher approves every mark.

01

Student writes

On paper, in class, as today

02

Teacher sends the class set

Upload, email from the scanner, or a link from an assistant

03

Pages are read

Handwriting turned into text

04

AI drafts marks

Against the teacher's own rubric

05

Teacher checks

Changes anything, then approves

06

Student sees it

Marks and feedback they can act on

Nothing reaches a student until a teacher has approved it.

Today

Every partner is its own code path.

A new partner means new code, a new release and weeks of engineering. Each one makes the next harder.

Seven Steps

Big Write VCOP

The Write Stuff

if partner ==

Today’s app: one branch per partner shape

Fixed standard

Chosen by the name "Seven Steps"

Criterion ladder

Chosen by the name "Big Write VCOP"

Write Stuff ladder, bespoke

Chosen by the name "The Write Stuff"

Why change

Each partner launch is slow, and the next is slower.

From signing to live
23 wk
Empowering Writers, signed 30 Apr, live 9 Oct 2026
Bespoke partner code paths
3
The Write Stuff added the third in November 2026
Files that check one partner's name
~40
In The Write Stuff's open pull requests
  1. MMW partner launch records and today's app code, read 13 November 2026.

Future

One platform. Partners are configuration.

Adding a partner is an import of rows, not a release.

MMW app for schools

Teachers and students

Partner platforms

Their brand in front

Publisher platforms

White-labelled, headless

the same API

Marking primitives, one API each

TranscribeMarkFeedbackValidate rubricMatch names

A partner is just rows:

RubricScore ruleModel adapterBrandSchool grants

Current and future

What changes for a partner launch.

Today

New code, a release and weeks of engineering.

Score rules hardcoded per partner.

Everyone must use MMW's screens.

Several rewrites in parallel with no shared plan.

New platform

Rows: rubric, score rule, model adapter, brand.

Data, checked by a validator the day it arrives.

Publishers call the same API from their own platform.

One repository, one plan, decisions written down.

Launching a partner

From months of code to a checklist.

The same partner, two ways. The new way needs no engineer and no release.

Today23 weeks for the last partner
01

Sign

Contract signed

02

Engineers code it

New branches for their rubric and scores

03

Release

Wait for the next release

04

Live

Schools start

New platformConfiguration, not code
01

Sign

Contract signed

02

Load their rows

Rubric, score rule, brand

03

Validator passes

Checked the day it arrives

04

School connects

School approves access

05

Live

Schools start

Why this shape

We copy platforms that already work.

Plus Stripe's habits: an idempotency key on every create, so nothing is marked twice, and webhooks that never carry student work.

Stripe Connect

They

A platform acts for many connected accounts, each with the capabilities it was granted.

We

A school connects a partner or publisher and approves what it may use: content, rosters, work.

Twilio

They

One API per capability (SMS, Voice, Verify) on one account with one set of keys.

We

Transcribe, mark and give feedback are separate APIs. A publisher uses any subset.

Resend

They

A small API with examples in every language. A developer sends their first email in minutes.

We

A partner goes live from our docs and their configuration, without an MMW engineer.

How a school connects a partner

The school stays in charge.

Like Stripe Connect: the partner asks, the school's admin approves each kind of access, and the API enforces it on every request.

Partner or publisher

In Australia, the UK or the US

asks

School admin approves

ContentRostersStudent workUsageStaff

Each one separately. Asking for more later needs a new approval.

allows

Partner can mark

Only for that school, only what was approved

How it fits together

One copy per region. Student work stays home.

Each region runs the same stack with its own database, files and models. A partner in any country calls the region its school lives in.

Shared across regions: which school lives where, adult sign-in, and partner and publisher content. No student work.
First build, Sydney

Australia

Its schools, and the partners and publishers serving them

API and web app

Every request scoped to one school

PostgresPagesML models

ST4S: today's platform assessed compliant, Medium risk (draft report). The new hosting is assessed next.

When a UK customer needs it

United Kingdom

Its schools, and the partners and publishers serving them

API and web app

Every request scoped to one school

PostgresPagesML models

UK GDPR. Its own assessment before the first school moves.

When a US customer needs it

United States

Its schools, and the partners and publishers serving them

API and web app

Every request scoped to one school

PostgresPagesML models

FERPA, COPPA and each state's student privacy agreement.

What never changes

Safe by design, not by policy.

Five rules every part of the platform is built and tested against.

1

A teacher approves every mark

Before any student sees it, and again if the AI's draft changes.

2

No machine releases marks

No agent, script or API key can approve or release.

3

Safeguarding comes first

A concern alerts the school's named contact and blocks release.

4

Student work stays in its region

And never trains a model.

5

Nothing is marked twice

Every change is recorded, and every request can safely be retried.

Who moves when

Australia first. Everyone else when their region is ready.

Nobody is switched off. Each group moves when the new platform can serve it, and today's platform runs until then.

Australian schools

Term 1 2027

New platform, Sydney

Partners and publishers

Set up before Term 1

Loaded as rows, then a two-month pilot

UK and US schools

When a customer needs it

New platform, their own region

One week in

The new system runs. Its marks are not real yet.

The goal set on 9 Nov was the new system running and the first publisher onboarding. This is where each stands.

Built since 9 Nov

The whole API

73 routes, running on staging in Sydney with made-up data only. Checked live on 16 Nov.

The school app's first screens

Teacher review and release, student sign-in, admin rosters. Tested in the code, not on staging yet.

A class set in one drop

Dropped as one batch, split by student, confirmed by the teacher.

Partner set-up as one file

One launch file per partner, checked and applied by one command. API keys and a sandbox.

Not real yet

The marking engine is a stand-in

It returns made-up text and marks until a real model is connected.

There is no production

Nothing holds real student work. No launch gate has a written pass condition.

Publisher onboarding has not started

The proof runs on made-up pages only. Next slide.

Code is no longer the slow part. Decisions and bookings are.

Publisher proof

The proof runs on made-up pages. Real pages are next.

Cambridge University Press asked whether we can mark whole handwritten pages with diagrams and tables. Our answer is due on 4 Dec.

Runs today on three made-up pages: a table, a diagram, a multi-part question

01

Worksheet as data

Question ids, answer boxes and the mark scheme

02

Find the answers

Stand-in: prepared regions, not real scans

03

Match to question ids

Unclear working goes to the teacher, never guessed

04

Mark each part

Stand-in marker, no model yet

05

Report

Accuracy, cost per page, failures by feature

Still needed for 4 Dec

Real pages

Due 20 Nov. Not received yet. We are checking two archives for maths pages and have asked for the rest.

Scope and a named builder

Not named yet. We ask James to agree the scope and the builder at this meeting.

A real model

Waits on the rule for outside AI models and a spend cap.

The plan

Everything built by the end of 2026.

Each step ends with something people can use or check, not a report.

  1. Nov 2026

    The API is built

    The whole API runs on Australian staging, on a stand-in engine.

  2. 4 Dec 2026

    Page-marking proof

    Whole pages matched to question ids for Cambridge.

  3. 11 Dec 2026

    Seven Steps parity

    Seven Steps marking matches today's platform on its gold set.

  4. 18 Dec 2026

    Launch gates

    Pen test, consent, safeguarding and restore drill all pass.

  5. 31 Dec 2026

    Done

    Built and checked. Australian schools move across for Term 1.

Risks

What could move the dates.

Each has a date or an action against it. None of them is code.

The team may be too small

Nothing yet shows the team can finish by 31 Dec.

What we do

The capacity estimate on 25 Nov. If it fails, the dates move.

Real pages may arrive late

The Cambridge proof needs real pages for 4 Dec.

What we do

If they have not arrived by 20 Nov, we tell Cambridge that day and move the date.

The launch gates are not ready

A security test of the new platform is to be booked this week. Consent and safeguarding are not specified.

What we do

Book a test of the new platform that reports before 18 Dec, and name the safeguarding lead, this week.

Real marking is not connected

Marks are made up until a real model runs in Sydney.

What we do

Ask for GPU capacity in Sydney now, and settle the rule on outside AI models.

If a gate is still open on 18 Dec, schools stay on today’s platform. Nobody is switched off.

Calls to make

What we need to decide next

ActionOwnerBy
  1. 01

    Capacity estimate

    Confirm the team can finish by 31 Dec

    Owner: Matthew, Callum

    By: 25 Nov

  2. 02

    Infrastructure owner

    Who runs the AWS accounts and the staging setup

    Owner: Matthew

    By: 25 Nov

  3. 03

    Outside AI models

    Self-hosted only for marking, or allow named cloud models

    Owner: Matthew

    By: Not set

  4. 04

    The residency rule, in writing

    Student work stays in its region. Real data waits on it

    Owner: James

    By: Overdue

From James

Decisions needed from James

ActionOwnerBy
  1. 01

    Agree the plan as the direction

    And say what the 4 Dec review needs to show

    Owner: James

    By: 16 Nov

  2. 02

    Safeguarding lead and security owner

    Named people. The safeguarding gate cannot be written without the lead

    Owner: James

    By: Overdue

  3. 03

    Publisher proof: scope and builder

    Who builds it, and who joins the meeting in the week of 23 Nov

    Owner: James, Matthew

    By: Overdue

  4. 04

    Spend caps

    One for staging and one for real model runs. Real marking waits on it

    Owner: James, Charlie

    By: Overdue

Your turn

Tell us what this plan gets wrong.

Send questions and objections to Matthew before 25 Nov.