About the Customer

‍

Univaciti is a Nigerian-founded EdTech platform established in 2019, serving students, early-career professionals, and corporate teams across online education and professional skills development. Its flagship programme, TESA (Tech Skills Accelerator), is an intensive 8-week bootcamp spanning Cloud Engineering, Data Analytics, Software Engineering, and AI Engineering. Additional offerings include STEM for young learners, TESA X and TESA PRO for corporate upskilling, and Certify for formal skills validation.

‍

Operating a hybrid virtual and on-premises delivery model, Univaciti positions itself as an outcome-driven platform with strong emphasis on hands-on training, industry mentor access, and pathways to hiring partners. As cohort sizes grew, Univaciti began investing in AI-powered capabilities to deliver personalised, intelligent learning experiences at scale — establishing itself as an emerging ISV and EdTech innovator on the African continent.

‍

‍

‍

Customer Challenge

‍

As Univaciti's cohort sizes scaled through 2024, three critical challenges emerged that threatened learner outcomes and operational sustainability.

‍

Personalised learning became impossible at scale. Every learner received the same static content regardless of pace or comprehension, with no mechanism for real-time intervention. In an 8-week intensive programme, a learner who falls behind without support is at high risk of dropping out — directly impacting revenue and platform reputation.

‍

Assessment could not keep pace with growth. The project-heavy curriculum created grading backlogs that stretched turnaround from same-day to multiple days, breaking the feedback loop essential to practical learning. Inconsistent grading across instructors added fairness concerns that exposed the platform to learner complaints.

‍

Without an intelligent support layer, learners who didn't understand course material had no recourse beyond waiting for instructor availability — an unsustainable model as enrolments grew.

‍

The platform's teams struggled with:

• No Personalisation at Scale: Static content delivery with no real-time intervention capability as cohorts grew
• Grading Bottlenecks: Multi-day turnaround on project-based assignments breaking the learning feedback loop
• Inconsistent Assessment: Grading quality varied across instructors, creating fairness and trust issues
• Instructor Burnout Risk: Growing volumes of routine support and grading requests displaced high-value teaching time
• Scalability Ceiling: The people-heavy delivery model could not support commercial growth without proportional headcount increases

‍

‍

‍

Solution Implementation

‍

In 2025, Univaciti partnered with Qucoon, an AWS Advanced Consulting Partner, to design and deploy Clara — a purpose-built agentic AI learning companion and intelligent assessment system running entirely on AWS. The engagement covered the full consulting lifecycle: defining use cases, architecting the solution, and delivering a production-ready implementation scaled to current learner volumes with defined growth triggers.

‍

Clara — Agentic Learning Companion:Clara maintains a persistent memory profile per learner through AWS AgentCore Memory, adapting responses to each student across sessions. When a learner asks a question, Clara queries Bedrock Knowledge Bases using hybrid vector and keyword search across Univaciti's actual course curriculum. A ReAct reasoning loop managed by AgentCore Runtime invokes a web search tool when course content alone is insufficient, with all external results sanitised by AgentCore Gateway before reaching the student. Responses are streamed in real time via Claude Sonnet on Amazon Bedrock.

‍

Two-Phase Intelligent Grading System:When an instructor creates an assignment, AWS Step Functions invokes the Grading Agent on ECS Fargate, which uses Claude Sonnet to extract a structured JSON score sheet from the rubric. Step Functions pauses at a human-in-the-loop gate, notifying the instructor via SNS and SES to review and approve the score sheet before any student can submit. On each submission, the Grading Agent grades criterion by criterion against the locked score sheet — with the student submission explicitly labelled as untrusted content to prevent prompt injection — and returns structured feedback to the learner within minutes.

‍

Content Pipeline and Question Generation:When instructors publish course content, an EventBridge event simultaneously triggers Lambda to call Claude Haiku for module summarisation and a Bedrock Knowledge Bases sync to index content into OpenSearch Serverless, building the RAG layer that grounds all Clara responses. A separate stateless flow generates and grades practice questions per module, also powered by Claude Haiku, with no persistent storage between calls.

‍

‍

‍

Key Technologies and Capabilities

‍

• Amazon Bedrock — Claude Sonnet: Deep reasoning for the Q&A agent and per-submission grading, with streaming and structured JSON output
• Amazon Bedrock — Claude Haiku: High-volume structured generation for module summarisation and question creation
• AWS AgentCore Runtime, Memory, and Gateway: Iteration budget enforcement, persistent learner profiles, and web search with prompt injection filtering
• Bedrock Knowledge Bases: Curriculum-grounded RAG using semantic chunking and hybrid vector plus BM25 retrieval
• AWS Step Functions: Two-phase grading orchestration with waitForTaskToken instructor approval gate and full audit trail
• AWS ECS Fargate: Warm persistent containers for the long-running Grading Agent
• AWS Lambda: Serverless execution for the Q&A agent, content pipeline, and question generation
• Amazon EventBridge: Decoupled event routing with Schema Registry across all pipeline consumers
• Amazon SQS: Grading queue with idempotency controls and Dead Letter Queue for failure recovery
• Amazon DynamoDB: Sub-millisecond key-value storage for sessions, score sheets, grades, and summaries
• Amazon OpenSearch Serverless: Vector search index backing the Knowledge Bases RAG layer
• Aurora PostgreSQL Serverless v2: Relational data layer that scales to zero during idle periods
• Amazon SNS, SES, X-Ray, and CloudWatch: Instructor notifications, distributed tracing, Bedrock invocation logging, and grading failure alerting

‍

‍

Results and Benefits

‍

The Clara platform, live in production on AWS, delivered measurable improvements across learner outcomes, instructor efficiency, and Univaciti's competitive positioning.

‍

Learner Outcomes:

• Above 80% Programme Completion Rate: Up from a 50–65% baseline prior to Clara, driven by real-time support at the exact moment learners are at risk of disengaging
• Grading Turnaround Reduced from Days to Minutes: Students receive criterion-by-criterion feedback within minutes of submission, closing the feedback loop while work is still actionable
• Persistent Personalisation: Clara's memory layer produces increasingly relevant responses as each learner's profile builds over the duration of their programme

‍

Operational Efficiency:

• Cohort Scalability Decoupled from Headcount: Cohort sizes can grow without proportional increases in instructor or support staff
• Consistent, Auditable Assessment: Every submission graded against a pre-approved score sheet, eliminating instructor variability and providing a defensible audit trail
• Instructor Focus Restored: Automation of routine Q&A and grading frees instructors for high-value mentorship and curriculum development

‍

Commercial Impact:

• Competitive Differentiation: Clara's agentic AI capabilities distinguish Univaciti from EdTech competitors entering the African market
• Strengthened Corporate Client Confidence: Consistent, auditable outcomes support enterprise clients who measure training effectiveness for their employees
• Live AWS Production Deployment: Running in production with an estimated ARR of approximately $1,476, with Amazon Bedrock accounting for approximately 30% of total AWS spend

‍

‍

‍

Lessons Learned

‍

The Clara implementation surfaced several insights that will shape future agentic AI deployments.

‍

Technical Lessons:

• Rubric Input Quality is an Architectural Constraint: Loosely written rubrics produced score sheets requiring significant instructor revision. Rubric templates and writing guidance were embedded directly into the assignment builder to standardise input before it reaches the AI — a step that should be designed into similar systems from the outset

• Human-in-the-Loop Gates Should Be Native, Not Added Later: Building the instructor approval gate into Step Functions from the start produced a cleaner, more auditable architecture and made human review feel integral to the workflow

• Prompt Injection Defence Must Be Structural: Labelling student submissions as explicitly untrusted within the grading agent's context window, as an architectural decision, proved more robust than relying on prompt-level instructions alone

‍

Operational Lessons:

• Confidence Indicators Build Instructor Trust Faster: Surfacing which criteria Clara was less certain about allowed instructors to review efficiently rather than reading full score sheet output from scratch
• Memory Write-Back Requires Careful Calibration: Writing to learner profiles only on meaningful comprehension signals — not every interaction — kept profiles accurate and prevented noise from degrading personalisation quality over time

‍

Strategic Lessons:

• Agentic AI Extends Experts, It Does Not Replace Them: The instructor approval gate is the clearest expression of this principle. Clara does the structured reasoning; the instructor validates the outcome. This framing drove adoption more than any technical capability
• Right-Size Compute to Workflow Requirements: Recognising early that the grading workflow required durable state and indefinite execution pause — and selecting Step Functions and Fargate accordingly — avoided a costly re-architecture later

‍

‍

‍

About the Partner

‍

Qucoon is an AWS Advanced Consulting Partner with deep expertise in cloud-native architecture, generative AI, and agentic AI solutions. Operating across Nigeria, Kenya, South Africa, the UAE, the UK, Canada, and the USA, Qucoon helps organisations design and deploy production-grade AI systems on AWS that are secure, scalable, and built for real business outcomes.

‍

Through engagements like the Univaciti Clara platform, Qucoon delivers end-to-end agentic AI consulting — from use case definition and architecture design through to production deployment and continuous improvement — empowering customers across Africa and beyond to harness the full capability of the AWS AI services stack.

‍