Capabilities · 24 named mechanisms
Grouped by the questions real buyers ask. Every item names the exact internal mechanism that powers it, a technical evaluator can verify rather than trust.
The architect + writer + translator pipeline that turns a brief into a course.
A reasoning-capable LLM proposes chapter + section structure + learning objectives. You approve or regenerate before the expensive content-writer pass fires. No wasted credits on a draft you'd reject. The architect model is configurable per organisation, bring your own LLM or use the Mill default.
Welcome, bullets, tabs, accordion, flashcards, MCQ, true/false, fill-in-the-blank, split-content, stat-cards, quote, image-showcase, hotspots, before-after, comparison-table, checklist, timeline, process-flow, decision-table, case-study, drag-drop, ranking, sortable, branching-scenario, reflection, numbered-steps, testimonial-carousel, key-takeaways, matching-pairs, word-cloud, video, image gallery, callout, definition, summary, and more. Every section has its own render template tuned for learning.
.docx, .pptx, legacy SCORM .zip, upload and Mill reads the text, extracts structure, and generates against it. Strict-source forbids claims not anchored in your source document. Defensible procurement-grade by default.
33+ languages, one generation pass. Translation runs through the configured translation engine with your per-org glossary (DO_NOT_TRANSLATE + PREFERRED_TERM). Narration renders through the configured TTS engine. Both are per-organisation: bring your own providers, or keep the Mill defaults. Generation cost for 10 languages is not 10×, it's ~1.3×.
The configured image model (per-organisation: hosted vendor, self-hosted diffusion, or the one your design team already uses) generates an image per section that asks for one. Capped (4-10 per course depending on complexity) so the course doesn't feel like a stock-photo collage.
Everything a procurement team asks about.
Every publish writes a `PublishVersion` row with SHA-256 manifest hash + model hashes. Years later the regulator can ask 'what did learner Y see on date X?' and we answer in one query, with content-addressable proof.
SME_REVIEW + COMPLIANCE_SIGNOFF signatures required before a course is 'fully signed'. Same user can't hold both roles. Reason codes (REVIEWED / APPROVED / AUTHORED / RESPONSIBILITY) + ipHash captured. Modelled on 21 CFR Part 11 §11.50/§11.70.
Translator output runs back-through-translation. Fields classified numeric / unit / regulatoryId / prose. Any divergent token on numeric/unit/regulatoryId = hard fail, regardless of semantic score. Stops the '5 mg → 5 ml' class of translation disaster.
AI-assistance banner renders on the authoring + learner surface. Model card drawer lists every AI provider + model + purpose. Flag-gated so legal can flip it off during a policy review without a redeploy.
All infra (app + DB + default AI provider calls) hosted in the EU. Each organisation can override per-role: pin to a specific LLM, route TTS through a contracted vendor, run image generation in your own VPC. Lock the model and skip the risk review.
What makes team plans actually productive, not just a seat count.
Team Starter: 3,000/month across 5 seats. Team Standard: 5,500 / 10. Team Pro: 9,000 / 25. No per-user quota; let the person who writes the most courses do it.
New team courses default private. The author flips 'Share with team' to expose it to their org. No accidental leaks to teammates who weren't in scope.
Your product names, brand terms, regulatory references, set them once, apply them to every generation + translation across every member.
Assign named reviewers per language. Status chips gate publish. Publishes block until the required reviewers approve.
Team Pro + Hall subscribers: Mill pushes published courses directly into the LMS. ORG_ADMIN pastes a token once; every future publish auto-imports.
Mill isn't fire-and-forget. The editor is where 90 %-good becomes ship-ready.
Every field of every section editable independently. Regenerate a single cue without touching the rest of the course. Per-language tabs for multilingual editing.
Inline OK / warn / fail pill next to each translated cue. Click to check; tooltip shows back-translation + divergence. Catches translator glitches before learners see them.
Edit → save → click Republish → a new SCORM/xAPI package is written + a new PublishVersion row is recorded. The learner opens the exact same URL and sees the new version next session.
Cache-Control: no-store on the preview response. Edit → republish → refresh preview. You always see the latest built bytes, never a browser-cached stale copy.
For the engineers who ask before procurement does.
Modern, maintained stack. EU-region hosting on Hetzner / Supabase or equivalent. Standard open-source tooling throughout, no proprietary lock-in.
Every AI role in the pipeline (architect LLM, content writer, translator + drift-check, TTS, images) is provider-agnostic. Mill ships with sensible defaults, but each organisation plugs in its own stack: a hosted frontier LLM, a self-hosted open-weights model, the TTS engine your voice team already licenses, a diffusion image model behind your VPC. Switch per-org or per-group without a code change. Audit log captures which provider + model served each call.
AI kill switches per provider, budget circuit breakers per-course + per-user-daily, Art. 50 banner toggle, SCORM storage backend toggle. Ops can flip any of these without a redeploy.
Every request carries an async-local correlation ID; logger emits JSON to stdout (picked up by Axiom / Better Stack / Datadog, your choice). One query pulls every AI call + DB write for a given course.
Mill org owner issues Bearer tokens scoped to their org. Hall + Zapier + custom scripts authenticate with Authorization: Bearer. Token hash stored; prefix kept plaintext for admin UI; revoke any time.
Seen enough?
130 free credits. No card. First course in the next quarter-hour.
