Behind the coaching conversation is a deliberate architecture. A purpose-built learning model, a global standards graph, a personalisation engine, and a measurement spine that turns every interaction into evidence. Here is what is under the hood, and why it is safe to put in front of your people.
Each layer does one job and hands off to the next. A learner's message travels down for grounding and up for delivery. Everything it touches emits evidence.
Three front ends on one backend: the learner app, the admin console and the embeddable Little Zav widget. Each one is white-labelled per client.
The invisible conductor. It reads a live flow-state signal each turn and chooses which teaching character responds. Challenge is tuned to ability in real time.
A roster of specialist coaches. Each is tuned to a level of Bloom's taxonomy, from building foundations to certifying competence.
A retrieval layer that grounds every turn in who the learner is: role, prior signals, neurodiversity preferences, goals. Responses fit the person, not the average.
The learning-science engine: a small, purpose-tuned model with a market-specific adapter for language, framework and cultural register.
601k qualifications and their skills, mapped across 150+ countries. This is the ground truth every lesson is built against.
Every meaningful interaction becomes an xAPI statement in a self-hosted Learning Record Store. That feeds reporting. It can feed your own LRS too.
Zavmo does not rent a general-purpose frontier LLM and hope it behaves. Saraswati is built on Microsoft Phi-3 Mini. It is small enough to run efficiently, and tuned for learning science. The MIT licence is commercial-friendly.
A model you tune is a model you can audit, test and constrain. It is not a black box whose behaviour changes when a vendor ships an update.
A lightweight LoRA adapter per market carries the local language register and qualification framework. UK is live. Ireland, Singapore and Australia follow the same pattern, with no retraining from scratch.
An SLM costs a fraction of a frontier LLM to run, per interaction. That is what makes one-to-one coaching affordable at organisational scale.
Subject knowledge is retrieved at run time from the standards graph and character corpora, not baked into weights. So content stays current and traceable.
The standards graph is a Neo4j knowledge base linking 601,000+ qualifications and their underlying skills across more than 150 countries and 30+ frameworks. It is the difference between generic content and content mapped to the exact skill your role and jurisdiction require.
National Occupational Standards define what good looks like for a role. They are the skills and behaviours a competent person demonstrates.
Regulated qualifications supply the learning content and assessment criteria. That is the structured curriculum that leads to a credential.
Zavmo maps a learner's role to the right skills, pulls the matching qualification content and generates a bespoke path. Grounded, not invented.
Before Saraswati generates a single word, the LearnerSignalBundle assembles what is known about the learner. Their role, prior interactions, stated goals, confidence and neurodiversity preferences. The response is grounded in that. Personalisation is not a greeting with a first name. It is the model reasoning from the real person.
The best way to learn has been known for forty years. The hard part was making it affordable for everyone.
The 2 Sigma problem, which this architecture exists to solveRather than one voice trying to do everything, Zavmo fields seven specialist teaching characters, each tuned to a level of Bloom's taxonomy. There is also a group-work mode where four team-mate avatars simulate a real workplace team. Agent 13 is the meta-agent above them. It reads a live flow-state signal each turn and picks the right specialist to keep the learner in the productive zone between boredom and overwhelm.
Patient mentor who lays solid foundations, step by step.
Hands-on practice partner with immediate feedback.
Precise thinker who surfaces patterns and connections.
Strategic guide who aligns learning to the career goal.
Zero-judgement zone to try things and learn from failure.
Four team-mate avatars, each a different workplace profile, for practising collaboration and judgement in a simulated team.
Chooses who responds, turn by turn, from the live flow signal. Full roster deck →
Learning that cannot be measured cannot be proven to the board. Zavmo emits xAPI statements for meaningful interactions, into a self-hosted yetanalytics lrsql Learning Record Store (Postgres-backed, Apache-2.0). xAPI is the open standard for learning data. It is wire-standard and vendor-neutral, so the same data can flow to your LRS too.
| Standard | xAPI 1.0.3 and 2.0.0. Portable, future-proof, not a proprietary format |
| Store | Self-hosted yetanalytics lrsql, backed by Postgres. One service, actively maintained |
| Forwarding | Vendor-agnostic wire format. Point statements at your own LRS via a single endpoint variable |
| Feeds | Admin reporting, learner insight (flow, learning personality), and market-level model improvement |
The features win the demo. This is the section that gets the deal signed. Zavmo is engineered against recognised standards from the start, not bolted on afterwards.
Engineered to the international information-security management standard. Access control, encryption and auditability are first-class requirements.
Built to the AI-management standard: governed model behaviour, documented decisions, and accountability for how the AI is trained and used.
Architecture supports EU data residency, so learner data stays where your compliance team needs it to.
Every query is scoped to the organisation. Admins see their people and no one else's. Isolation is enforced in the data layer, not just the UI.
Model improvement uses aggregated, market-level learning signals. Never individual records. Character knowledge is retrieved (RAG), not baked into weights.
Responses are grounded in the standards graph and quality-scored. The admin Prompt Studio lets teams test prompt changes in simulation before they reach a learner.
| Identity | SSO for learners and admins. Zavmo slots into your existing identity provider |
| Content standards | SCORM and xAPI interoperability for existing LMS/LXP estates |
| Data out | xAPI statement forwarding to your Learning Record Store or data warehouse |
| Embedding | Little Zav drops into a host LMS as a widget. Full white-label theming per client |
| Delivery | Cloud-hosted, continuously deployed, with a protected release pipeline and automated quality gates on every change |
A purpose-built model, a real standards foundation, isolation and governance by design, and open, portable measurement. If your security, data or L&D team has questions, we would rather answer them in detail than gloss over them.
Book a technical deep-dive