Loading
Every feature we're actively working on, with rough ETAs where honest. Plus: what we've decided NOT to build, and the specific reason each was cut. Public, dated, updated when scope moves.
If it’s on our pricing page, it’s live today. Items here are on the workbench — and the ones marked shipped came off it recently and are live now. If something is neither, we haven’t decided yet.
Draw your working in the pad, or photograph the page, and it is transcribed into the attempt box for you to correct before you send it. The transcriber only transcribes: it will not finish an unfinished line or fix a wrong one, because the mistake is the part the tutor needs. Next on this thread: diagrams read as diagrams rather than described.
Hold to speak your explanation and the browser transcribes it. The transcript lands in the box for you to correct before it's graded, because speech recognition mishears technical terms and this answer counts. Browsers without the API get the typing path.
Second Supabase region for institutional customers with hard EU-residency requirements. All new EU customers post-launch will default to the Frankfurt region. Existing customers can migrate on request.
Formal SOC 2 Type I attestation. Not started — it is queued, not under way. What exists today is our own threat model, an application self-audit and the remediation list from it, which we will share with institutional prospects under NDA. No third party has penetration-tested us yet.
Opt-in cohort view for classmates. Aggregate progress only — never each other's attempts. Educators can already do this via /for-schools; this brings it to peer groups.
Paste a link to a public course page or reading and it's fetched, chunked and indexed like any upload. Alongside it: PDF, Word, PowerPoint, Excel, LibreOffice, EPUB, photos of slides and notes, CSV, notebooks, and SRT/VTT lecture transcripts — many at once.
The web app is already installable as a PWA. Native wrappers with push notifications for recall reminders land next. Feature parity from day one.
Every problem, hint and recall card is generated from the syllabus, slides, notes and readings you upload, and cites the page or slide it came from. There is no generic bank. Next on this thread: targeting the generator at your weakest concept automatically.
Friday recap: what you learned, what you struggled with, what to attack next week. Optional shareable card. Different from the Sunday Brain-score email.
A gentler, larger-type, image-heavy variant of Learn mode for elementary. Requires a different content-safety layer + parent co-pilot mode.
Grammar + conjugation practice is already good in Learn mode. What's missing is the vocabulary-acquisition surface (which Anki does well). Deciding whether to build it or partner.
Multi-student real-time attempts on the same problem. Could be great; could be a distraction. Prototyping to see whether it protects the attempt-first workflow.
Every quarter we get emails asking us to build these. We don’t — not because they wouldn’t sell, but because each one would collapse a specific load-bearing part of the product. The refusal to build is as public as the roadmap.
why · It would collapse the refusal into a soft answer machine. Not for sale at any price.
why · Even with a 'don't submit this' checkbox, we'd be publishing the exact material we exist to withhold.
why · The ladder's cost is what makes hints intentional. Bypassing it destroys the mechanism.
why · There will never be ads on Brainback. The business model doesn't require it.
why · Even anonymized, it's a category of use that trains us to think about data as inventory. We won't.
why · See /parents/privacy. Content-blind is load-bearing, not a preference we'd flip for a bigger deal.
Nothing here is a promise. Ship dates in tech are rough estimates that slip; we publish them because slipping estimates are more useful than no estimates. When an item moves — up in priority, down in priority, or out entirely — we update this page and the changelog notes the move.
What’s worth trusting: the “won’t build” list is more stable than the “will build” one. Every item in the former has been rejected multiple times.
A real human reads every feature request. If it’s a good one, we’ll say why we agree; if we disagree, we’ll say why we disagree. Either reply within 24 hours.
No auto-reply. A human reads every one.