Everything works offline
Courses, assessments and credential issuance all run against local storage. A connection is an optimisation, not a requirement.
Technology
Alexandria is one native codebase that turns a device into a full participant in a learning network — issuing credentials, assessing skills and connecting peers, with no server in the middle that has to be trusted. Here is how each part of that works.
01 / Identity & credentials
Launching the app generates a keypair locally. That key is your identity — there is no signup, no email, no password and no record of you on any server of ours. It signs every credential you earn and encrypts everything you keep.
Each credential is a W3C Verifiable Credential signed with Ed25519 under your own DID. Six kinds of credential exist, and each is a separate object you can share on its own.
A hash of the evidence is also intended to be anchored on-chain, so a credential’s existence and timing can be proven independently of us. That part is in development — a credential stands on its signature without it, which is why the verifier on this site reports the anchor check as skipped rather than passed.
// issued on device, no server involved
{
"@context": ["https://www.w3.org/ns/credentials/v2"],
"type": ["VerifiableCredential", "SkillCredential"],
"issuer": "did:key:z6MkhaXgBZD…QYtP",
"credentialSubject": {
"skill": "welding.pipe.6g",
"level": "apply",
"evidence": "blake3:9f2c…e17a"
},
"proof": {
"type": "Ed25519Signature2020"
}
}02 / Data & sync
Courses, assessments and credential issuance all run against local storage. A connection is an optimisation, not a requirement.
Your laptop and your phone sync directly, end-to-end encrypted. Relays exist to pass sealed bytes along and cannot read a thing.
There is no server-side record of what you study, how fast you learn or how often you fail. Not policy — architecture.
03 / Sentinel
A credential is only worth what the assessment behind it is worth. Sentinel is an optional integrity layer that watches for the obvious ways an assessment gets faked — and runs entirely on the learner’s own hardware.
It takes quiet snapshots every 15–45 seconds and scores consistency across typing rhythm, pointer behaviour and small on-device models. Raw camera, keystroke and pointer data never leave the machine; what leaves is a score and the reasoning behind it, attached to the credential for anyone to audit.
Every proctoring vendor asks you to trust a black box. This one is inspectable: the code is open, and the signals that produced a score travel with the score.
0.94
Integrity score · High
Raw camera, typing and pointer data never leave this device.
04 / Assessment plugins
Different skills need different ways to teach and test. A plugin is a self-contained assessment engine — it defines what a candidate does, what evidence is captured and how it is scored — and once installed it can issue credentials like any other assessment. They run as sandboxed WebAssembly, so a plugin cannot reach anything the learner did not give it.
A real editor with a test harness. Scored on what runs, not on what is recalled.
Audio capture and pitch analysis. Play the passage; the plugin hears whether you played it.
Structured evidence capture plus assessor attestation, for welding, inspection and site work.
The plugin API is open. If you can define how a skill is proven, you can ship the assessment for it.
05 / The skill map
Every lesson and every credential ties to a node on a public map of skills, with explicit prerequisites and Bloom levels — recall, apply, analyse, and beyond. That is what lets an employer read welding.pipe.6g · apply and know exactly what was demonstrated.
The map is maintained by the people who hold the skills, not by us. Changes go draft → committee → public vote, and only people with proven expertise in that subject get a vote on it. Not money, not seniority.
Dot size shows level. Filled = Apply or above.
06 / People, not servers
Video, audio and screen-share connect learner and mentor directly. No call runs through a company, and none is recorded by one.
A teacher’s own contained space, organised into channels — announcements, questions, assignments, showcase. You decide who is in it.
Instructors are scored on measured impact on learners, per skill, with confidence bounds — wider bars mean less evidence. There is no global score anywhere.
07 / Runs everywhere
For live video tutoring on Android we recommend a device with 6 GB or more of RAM. Nine languages at launch, including Hindi, Bengali, Telugu, Marathi and Urdu.
Honest limits
All of the above is built and shipping in the alpha, apart from on-chain anchoring, which is in development. None of it has been through an independent validity study, there is no content moderation in place, and no external users have stress-tested any of it. The security audit is in progress and we will publish its findings, open or fixed.