LUMINA — Global Classroom Intelligence System
Automated, board-aligned IEP generation across more than thirty international SEN boards and two hundred subjects in fifty-eight languages — including full RTL — with sub-three-second median latency.
Context
The Individualised Education Plan is the single most consequential document in a special-needs learner's education. It defines accommodations, the assessment trajectory, the therapy hours, the resource allocation, the goal cadence, and the family-school contract. In most jurisdictions, an IEP takes three to six weeks to produce by hand. In many jurisdictions, the document arrives partial on the board-alignment side because the teacher writing it does not have time to cross-check the specification of every relevant SEN authority.
Challenge
Building an IEP generation system that satisfies thirty-plus SEN boards is not a single-schema problem. The UK SEN code of practice, the US IDEA framework, the Australian Disability Standards for Education, the Dutch Speciaal Onderwijs framework, the Pakistani LSU framework, the Indian RPwD framework, the UAE School For All framework, and the Canadian provincial IEP frameworks each define different terminology, different goal categories, different review cadences, and different stakeholder signatures. A single canonical schema cannot honour all of them. The path forward had to be a per-board translation layer over a structured intermediate representation.
Approach
The platform runs a RAG layer over a curated curriculum corpus of board specifications, subject content, and SEN diagnostic frameworks. Each board carries its own retrieval key namespace and its own citation-tracking layer, so that the generated document can be traced back to the specific clauses of the source authority. The IEP generation engine emits a structured JSON document validated against a per-board JSON schema; the document then renders to PDF, DOCX, or HTML through a templating layer keyed to the board's preferred document conventions.
Server-sent events stream the IEP back to the teacher's surface as it generates, so the wait time feels conversational rather than batch. The teacher can regenerate the document against an updated diagnostic in under three seconds, which means the IEP stays current with the learner across the review cycle rather than freezing at the original assessment.
Delivery Ownership
This engagement followed the Phoenix Group principal-led model. I owned the architecture, the curriculum corpus curation, the per-board schema design, the citation-tracking layer, and the document templating engine end to end. The SEN credential layer underneath the design — the experience certificates from two Karachi SEN institutions, the RBT 40-hour certification, and the active CBT diploma cohort — is what made the schema design work, because the engineering decisions had to honour clinical and pedagogical realities I had personally taught against.
Outcomes
Technology Stack
Next.js 15, FastAPI, Anthropic Claude API as the primary reasoning model, LangChain for the RAG layer, pgvector for the storage tier, PostgreSQL with Row Level Security for the per-tenant data plane, server-sent events for the IEP streaming flow, a per-board JSON schema validator, a board-keyed templating engine for PDF, DOCX, and HTML emission.
Services Delivered
RAG & Knowledge Fabric / Agentic AI Architecture & Build / LLMOps & Production AI Platform.
Why it matters
The IEP is the artefact a parent of a special-needs child reads more carefully than any other school document. The previous reality is that the IEP arrived weeks late, partial, and disconnected from the actual diagnostic. LUMINA changes that arithmetic. The teacher can produce a complete board-aligned IEP in under three seconds, regenerate it against an updated assessment whenever the learner's profile changes, and hand the family a document that cites its sources. For the schema-design detail behind the per-board translation layer, see the companion Insight: IEP generation across thirty SEN boards: the schema we ended up with.