Technology
An architecture designed for evidence and scale
ARASH separates presentation, application, AI services, business logic, data and infrastructure with strict one-way boundaries — so engines can change without touching the classroom experience.
Principles
Six engineering commitments
Six-layer architecture
Presentation, Application, AI Services, Business Logic, Data and Infrastructure with strict one-way boundaries.
Vendor-neutral AI gateway
No provider SDK exists outside the gateway, so engines can be swapped or degraded without touching product code.
Version-stamped scoring
Every score carries its rubric, prompt and model version, making results reproducible months later.
Domain-isolated data
Core, Learning, Conversation, Assessment, Analytics and Administration domains with row-level security per tenant.
Degradation ladder
If an engine is unavailable, the platform reduces depth instead of failing the learner's session.
Scale targets
Designed for 10,000 concurrent learners with 99.9% availability on the learning path.
AI services layer
Specialised engines, one gateway
Each engine has a single job and a versioned contract. No vendor SDK exists outside the gateway.
- 1
Speech Recognition AI
Transcript, word timings, confidence
- 2
Conversation AI
Scaffolded dialogue turns and hints
- 3
Pronunciation AI
Phoneme accuracy, stress, intonation
- 4
Grammar AI
Error taxonomy, severity, correction
- 5
Vocabulary AI
Range, CEFR band, target-word coverage
- 6
Fluency AI
Rate, pauses, disfluency, mean length of run
- 7
Comprehension AI
Relevance, accuracy, instruction following
- 8
Pragmatic AI
Register, politeness, turn-taking
- 9
Feedback AI
Student, teacher and parent feedback variants
- 10
Adaptive Recommendation AI
Mastery update and next-activity plan
- 11
Learning Analytics AI
Trajectories, cohorts, at-risk signals
System layers
Six layers, strict boundaries
Presentation Layer
Role-aware clients, speech capture, realtime rendering
Application Layer
Gateway, identity, orchestration, quotas, contracts
AI Services Layer
AI gateway plus specialized engines and rubric aggregation
Business Logic Layer
Domain invariants, pedagogy policy, workflows, events
Data Layer
OLTP, OLAP, object storage, cache, research vault
Infrastructure Layer
Cloud, edge, queues, secrets, observability, DR
Backend concerns
Ten disciplined concerns
Each concern has one owner and one place to change.
Controller
Transport handlers, DTO mapping, response envelopes
Middleware
Trace, authN, tenancy, authZ, validation, rate limit
Service Layer
Use-case orchestration and transaction boundaries
Repository Layer
Aggregate persistence, unit of work, row-level tenancy
API
Versioned public, private, AI, admin, analytics, research surfaces
Authentication
Tokens, sessions, SSO, guardian consent gating
Caching
Edge, HTTP, distributed, AI response cache
Logging
Application, audit, security, AI interaction streams
Queue
Scoring, analytics, notify, export lanes with DLQ
Storage
Private audio buckets, signed URLs, lifecycle tiers
API surfaces
Nine versioned surfaces
Separated by audience so security, quotas and deprecation stay independent.
Talk architecture with us
Happy to walk your IT team through security, retention and integration.