Platform reveal

What is new on MMW

Everything new on the platform, one reveal at a time, with links to where it lives.

Presented by
Matthew Blode
Date
November 2026

New branding

A new look is on its way.

Yoav is building the new look as living brand guidelines.

In progressOwner: Yoav

  • The brand studioOne typed brand file drives every swatch, WCAG contrast grade and type sample, and lists where the repo disagrees and what is still pending. Yoav owns it.
  • What we use todayCalm, plain and specific. The AI drafts, the teacher decides. Blue is the app's primary colour; the rest of the identity is still provisional.

Screenshot of the brand studio goes here.

The brand studioapps/mmw-design/apps/brandingThe brand guide todayapps/web/BRAND.md

The studio is not deployed and has no sign-in. Run it with bun run --filter @mmw/branding dev, then open http://localhost:3400/mmw.

New API docs

The API has its own docs site.

Partners and publishers learn the API from one public site.

Built

  • Six guidesIntroduction, quickstart, authentication, error codes, versioning and a changelog.
  • An API referenceBuilt from the committed OpenAPI document, so it matches the API.
  • The safety rule, shown in codeThe quickstart makes a marking, then shows that an API key cannot approve a mark.
  • Published by a workflowA workflow publishes the site when the API's contract or the guides change.
The docs sitemmw-docs.apidocumentation.com/Where the guides livescalar-docsThe quickstartscalar-docs/quickstart.md

The docs describe the API only and hold no student data. The API behind them runs on staging, not in production.

The school portal

Seven steps from a class set to a released mark.

One frame around every teacher screen, and one path through it.

Built on main

  1. 1Sign inHome says how many submissions wait for review.
  2. 2Drop a class setDrop the scanner's PDF on Class sets. No assignment needed.
  3. 3Check the splitMMW sorts the pages by student. The teacher confirms who is who.
  4. 4Open an assignmentIts progress sits in one line: marking, waiting, approved, released.
  5. 5Start reviewOne button on Home opens the first submission in the queue.
  6. 6Read and correctQueue, pages and marks sit side by side. A safeguarding flag sits above the pages.
  7. 7Approve, then releaseThe teacher approves the marks, then releases them one submission at a time.
  8. Screenshots of Home, an assignment and the review screen go here.

Runs on a laptop with seeded school data. The web app is not deployed yet, and the screenshots are the last open task of the change.

The change and its specopenspec/changes/add-school-portalThe screensapps/web/app/teach

A maths marking example

A maths worksheet, marked.

A worksheet goes in on paper and comes out as drafted marks for a teacher to check.

Built, runs locally

  • The walkthroughA walkthrough takes a teacher through a Year 8 worksheet in the seeded sandbox school. Every page, question and answer is invented.
  • What it showsThe AI drafts the marks against the teacher's rubric. The teacher approves every one.
  • Why a stand-inThe fake marker keeps the whole flow testable without a real model or real student work.

Screenshots of the maths worksheet and its marks go here.

The maths walkthroughdocs/demo/maths-walkthrough.mdThe fake markerml/page_marking

Marks in this example come from the fake marker, a stand-in for the real model.

The review stamp

A teacher stamps every mark.

The teacher's approval is what lets a mark reach a student.

Built

  • No stamp, no releaseA teacher approves every mark before a student sees it.
  • Tied to one version of the marksThe approval covers the marks the screen showed. If they changed meanwhile, the screen says so and approves nothing.
  • Never a machineNo agent, script or API key can approve or release marks.
The safety rulesAGENTS.mdThe review screen's requirementopenspec/changes/add-school-portal/specs/school-portal/spec.mdThe approval testapps/web/e2e/tests/review-approve.test.ts

Safety rules 1 and 2 in AGENTS.md. No change may weaken them.

Datadog

Datadog watches for trouble.

Datadog is where we watch the API and the web app for errors and slow pages.

Built, not yet checked on staging

  • The web appBrowser sessions go to Datadog's AP2 (Sydney) site, with all text and inputs masked.
  • The APIThe API can send traces once switched on. A filter drops SQL, URLs and error text before anything leaves.
  • Off until configuredBoth stay off until a setting names where to send, so a laptop sends nothing.
The observability packagepackages/observability/README.mdHow the web app reportsapps/web/AGENTS.md

Built in code. Nobody has yet checked that traces and alerts arrive from staging, so this deck does not count them as working.

PostHog

PostHog counts how the product is used.

PostHog counts what happens in the product. Events carry ids, so they are pseudonymous, not anonymous.

Built, not switched on

  • Pseudonymous, not anonymousBrowser events keep a browser id, a device id, a session id, the school and the role. Never names, emails, student work, prompts or error messages.
  • Nothing on student pagesThe student pages start no telemetry and do not even download its code.
  • Production only on the serverThe API sends status, duration and operation names from production only, and only with a host and key set.
The privacy rulespackages/observability/README.md

Built in code and off until a host and key are set. Nothing is sent today.

AWS Sydney staging

The API runs in AWS Sydney on staging.

Staging is in Sydney, built so student work stays in Australia.

Staging onlyOwner: Richie

  • What runs thereThe API on Fargate, Postgres on RDS and S3 for pages, all in ap-southeast-2.
  • How it gets thereA merge to main deploys once its checks pass. A failed migration stops the deploy.
  • What is not there yetThe malware scan is off on this account for now, and database backups are kept for one day.
The staging setupinfra/README.mdThe production changeopenspec/changes/add-production-environment

Staging holds synthetic data only. Production does not exist yet.

Safer student data collection

Safer ways to collect student work.

A teacher drops a class set and the system sorts it out, without student work passing through an assistant.

In progressOwner: Matthew

  • A safer uploadFiles go straight to private storage in Sydney, bound to their type, size and checksum. A malware scan is meant to check them before anything reads them.
  • Three ways inUpload in the app, email from the school's scanner, or a link started from Claude or ChatGPT. One step sorts the pages by student.
  • The teacher decidesThe teacher confirms the split. Nothing is marked until they do.
The changeopenspec/changes/add-intakeIts proposalopenspec/changes/add-intake/proposal.md

GuardDuty malware scanning is off on staging until the AWS account moves off the Free plan. Today a stand-in only looks for the EICAR test string. Also open: typed answers (waiting on a product question) and connecting Claude and ChatGPT on staging.

What's next

Target dates to the end of 2026.

Still open from today: branding (Yoav), AWS production (Richie) and safer student data collection (Matthew).

  1. Now

    Built, not in production

    The API runs on AU staging, and the school portal is built on main.

  2. 4 Dec 2026

    Page-marking proof

    Whole pages matched to question ids for Cambridge.

  3. 11 Dec 2026

    Real marking

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

  4. 18 Dec 2026

    Launch gates

    Production exists, and the four launch gates pass.

  5. 31 Dec 2026

    Done

    Built and checked. Then a pilot in Term 1 2027.