In betaFive more products, out of the same harness.
Amarna Studio proved the loop: a model writes a small structured document, gates check it before anything downstream sees it, and a deterministic compiler turns it into the real artifact. That loop is not about e-learning — it's about never letting a model hold the thing you ship. These five point it at five different problems, and the last two show its other half: when what you ship is an answer or a score rather than a file, the gates move to the moment it goes out.
All five are in beta: moving quickly, open to early users who want a say in where they go next. Each card ends with the line saying where that one’s edge is today — on this shelf those are the sentences that matter most.
Amarna Video
BetaType a premise. Get an animated episode you can actually edit.
// premise → document → animated episode
Everyone else generates pixels. Amarna Video generates a document that renders. A model writes ten fields per beat — actor, framing, action, emotion, line — and a pure, seeded compiler synthesises every bone transform, keyframe, mouth shape, particle and camera move from it. Nothing is sampled at render time, so nothing drifts: rewrite one line and the cast, the sets, the blocking and the shot list come back byte-identical, with only that line's mouth track and the timing after it moved. That addressable interior is the thing generative video structurally cannot have.
- Restyle without regenerating — 12 style packs over one document
- 54 animation controls saved as data, so a look is a link
- Exports mp4 to 2160p in the browser, with no API key required
- Style packs
- 12
- Animation controls
- 54
- Writing registers
- 6
Runs entirely in the browser today, keyless. Export quality follows WebCodecs support and the app tells you which path it took.
Amarna Agent
BetaAn agent that makes your phone calls.
// your week → a mandate → a call that gets made
Sign in with Google and your week shows up. Point at any line of it and say what the call is for — "ask if the brakes are done" — then hand it off. Amarna dials the business, talks to a human, and comes back with a transcript, a recording and whatever it agreed to on your behalf. What bounds it is a written mandate on every task: what it may agree to, what it must not, and the facts it is allowed to give out. Outside that list it says it needs to check with you and asks for a callback. It does not improvise.
- Reads the calendar and drafts both the task and its mandate
- Calls businesses, never people — enforced in the schema, not a policy page
- Every call opens by disclosing who it is for, that it is an AI, and that it is recorded
- From task to dialing
- 1 tap
- Independent business-only checks
- 3×
- Calls that disclose
- 100%
Places live calls through a provisioned number. Google Calendar is a sensitive OAuth scope, so early access is by invitation while verification completes.
Amarna Dental
BetaThree revenue apps for the practices nobody built software for.
// three apps · one platform · run one or run all
A $1m practice leaking 8% to denials and underpayments loses $80k a year, and almost none of it is visible by eye — an underpayment posts as an ordinary write-off and the ledger still balances. Recover reads the 835 remittance advice and turns X12 adjustment codes into worked appeals. FeeIQ ranks every payer by allowed production per chair hour and solves for the assumption a drop-or-renegotiate decision actually rests on. Pulse decomposes a change in collections into contributions that sum exactly to the total. All three are administrative, so FDA-exempt, and all three read Open Dental first — no five-figure integration toll.
- Recover — denial and underpayment recovery, with scoring you can argue with
- FeeIQ — payer scorecards, a break-even solver and a negotiation packet
- Pulse — benchmarking with attribution that adds up, no residual
- Apps, independently runnable
- 3
- Tests over the domain logic
- 61
- Per month, per practice
- $249+
The domain logic is real and tested, running today against a seeded synthetic practice. Live Open Dental integration and the BAAs land with the first design partners.
Amarna CX
BetaThe AI customer service department, measured honestly.
// question → grounded answer → a resolution someone checked
Every support platform sells deflection — the share of customers who stopped talking to you. It scores the ones you answered wrongly and the ones who gave up as wins, which is how a vendor shows 90% deflection against 40% real resolution. Amarna CX does not compute it anywhere. Sending an answer only claims a resolution; after the re-contact window a sweep asks the one question that settles it — did this customer come back, in this thread or any other — and marks it verified or reopened. Every rate is read off that ledger rather than a counter, because a counter can be nudged and a ledger has to be rewritten. Grounding is held to the same standard: every price, percentage and duration in an answer must appear literally in a retrieved passage, or the answer does not go out.
- Tenant isolation is a database invariant — a query that forgets its scope returns nothing, not somebody else’s transcripts
- Autonomy is granted one intent at a time, never by default
- Runs end to end with no API keys at all, on built-in simulators that say so on every screen
- Escalation triggers, worst case first
- 9
- Tests, isolation included
- 72
- Deflection metrics
- 0
Web chat is the channel that works today; voice, SMS and email are typed seams that refuse rather than pretend. Row-level security is inert against a Postgres superuser, so it ships a non-bypassing role and refuses to start without it.
Amarna LMS
BetaThe one LMS that knows what the course was trying to teach.
// course spec → evidence → proficiency, objective by objective
Every other LMS receives a sealed SCORM package. It can report that somebody opened it, spent 22 minutes and scored 80% — and nothing at all about what that person can now do, because the content is opaque to it. Amarna owns both sides of the seam: a course authored in Studio arrives here as a competency graph, with objectives and a quality, a score delta and the objectives it touches on every choice and every interaction. So the readout is per-objective proficiency rather than completions. Objectives with no evidence behind them are reported as exactly that and counted against the roll-up — quietly dropping them is how a report claims full coverage of a course that never assessed half of what it promised.
- An xAPI profile, not a seventh publish target — the same package still runs in Docebo
- Courses authored elsewhere are accepted too, and labelled imported rather than refused
- Refusals say “not found”, never “not yours” — ids are guessable and the difference is an oracle
- Tests over the domain logic
- 73
- Tables in its own schema
- 10
- Dependency-free domain modules
- 4
The furthest back of the five, and the only one with no running surface yet: the domain modules, the schema and the deposit contract are real and tested, but a deposit is validated today rather than stored. The proficiency thresholds are deliberately unresolved until the first design partner.