RJBack to selected work

Private client system

BinaAI

A curriculum-aware AI exam preparation product built around grounded answers, citations, structured assessments and controlled access.

Disclosure
Sanitized public case study
Role
Backend / AI Engineering
Stack
Django · FastAPI · PostgreSQL · React
Access
Private client project
About this case study

An overview of the architecture and engineering decisions behind a learning platform serving 1,200 students, with 60+ ingested textbooks and 500+ hours of video content.

The problem

Useful AI had to stay inside the curriculum.

An educational assistant cannot simply answer broadly and sound confident. The system needed to understand whether a request belonged to the supported learning scope, retrieve from approved material, return grounded responses and fit those answers into an exam-preparation workflow.

Curriculum-aware question handlingGrounded responses with citationsQuiz and assessment generationStructured model outputsTrial and usage gatingDocument-backed retrieval

Product flow

01Student request

Question or study action enters the product

02Scope classification

Determine whether the request belongs inside the supported curriculum

03Retrieval

Find relevant material from approved educational content

04Generation

Produce a structured answer or assessment

05Guardrails

Keep output within the learning workflow and supported scope

06Persistence

Store product state, usage and learning interactions

What mattered technically

Product boundaries were part of the AI architecture.

Scope before confidence

The product first needed to decide whether a question belonged to the supported curriculum. That boundary is more important than simply generating a plausible answer.

Ground the learning experience

Retrieval and citations help turn the model from a generic conversational layer into an assistant attached to the material the student is expected to learn.

Structured outputs for product behavior

Quizzes, assessments and workflow actions benefit from predictable schemas rather than free-form text. Model output becomes easier to validate and integrate into the UI.

AI features still need normal backend discipline

Authentication, usage limits, persistence, API boundaries and service ownership remain core product concerns even when the most visible feature is an LLM-powered assistant.

Application

Django REST · FastAPI services · React

Data

PostgreSQL · document storage · vector retrieval

AI product patterns

Scope classification · citations · structured generation · gated usage