2–3 Hours → ~3.5 Min
AI Assessment Creation & Rubric-Aligned Research-Paper Evaluation
How Cogninest AI built an agentic assessment engine on AWS and LangGraph, cutting grading turnaround from hours to minutes while grounding every mark in verbatim evidence.
Manual question drafting, evaluation bottlenecks, and feedback delays stalled student growth
Educational institutions and tutors face persistent operational bottlenecks in both assessment preparation and grading. Converting textbooks, syllabi, and educational materials into rigorous, balanced assessment items previously required hours of manual drafting, with question quality and depth varying widely across educators and departments. For long-form student research papers, grading against complex, multi-section institutional rubrics consumed 2 to 3 hours of tutor time per paper. Because feedback was typically delivered only after final marks were submitted, students lacked opportunities to revise their work iteratively. Generic AI chat tools proved inadequate—frequently delivering hallucinated or superficial comments without systematically mapping each mark to institutional rubric criteria and verbatim textual evidence from student submissions. Without strict orchestration, uncontrolled model inference also threatened institutional budgets as adoption scaled.
- Research-paper grading against detailed rubrics required 2–3 hours of tutor time per paper, creating evaluation backlogs.
- Converting raw learning materials into rigorous, balanced assessment questions consumed hours of educator time.
- Students received evaluation feedback only after final grading, preventing iterative learning and revision cycles.
- Generic AI tools produced broad commentary lacking verifiable citations from the student paper mapped to specific rubric mark bands.
- Unpredictable model token consumption risked escalating inference costs as submission volumes grew.
Dual-Engine Assessment Architecture: Bedrock Question Engine + LangGraph Generator–Reflector Pipeline
Cogninest AI architected and delivered a production-grade, two-tier AI assessment platform for BakPak. The system pairs an Amazon Bedrock-powered question generation workflow with a specialized multi-agent evaluation engine orchestrated by LangGraph and DeepAgents, supported by asynchronous background workers and real-time WebSocket updates.
Phase 1: Rubric & Multimodal Document Ingestion
Engineered automated ingestion pipelines for educator learning materials (PDF, Word, text) and institutional Excel rubrics. Leveraged PyMuPDF to parse text, tables, equations, and figures from student submissions into structured evaluation inputs.
Phase 2: Automated Question Generation on Amazon Bedrock
Built an educator-supervised question generation pipeline using Anthropic Claude on Amazon Bedrock. Enables educators to generate multiple-choice, short-answer, and long-form questions with granular customization and approval before distribution via secure links.
Phase 3: Generator–Reflector Agentic Grading Engine
Constructed a dual-agent LangGraph evaluation pipeline. A generator agent evaluates student papers section-by-section against the rubric and quotes supporting evidence directly from the text; an independent reflector agent audits each score and rationale before finalization.
Phase 4: Asynchronous Parallel Processing & Real-Time Delivery
Deployed Celery workers and Redis queues for parallel paper processing, paired with Django Channels WebSockets for live status updates. Students receive instant, tiered (brief or comprehensive) feedback to power iterative revision.
Phase 5: Observability, Usage Governance & Secure S3 Storage
Integrated Langfuse for end-to-end tracing of model latency, token consumption, and cost attribution. Implemented institutional submission quotas and secured all student papers and evaluation artifacts in Amazon S3.
"The BakPak assessment platform not only cuts research-paper grading from hours to minutes, but ensures every single mark is backed by transparent, rubric-aligned textual evidence. It gives our educators valuable hours back while providing students with actionable feedback in time to actually learn and revise."
Platform Leadership
BakPak AI Learning & Assessment Platform
Transforming Academic Assessment Through Speed, Consistency & Verifiable Evidence
The platform converted grading from a manual, high-fatigue bottleneck into an automated, consistent, and fully auditable workflow—empowering educators and giving students immediate, actionable guidance.
~3.5 Min
average grading turnaround per multi-section research paper
97%↓
reduction in educator grading time compared to 2–3 hour manual baseline
100%
marks backed by rubric criteria and quoted student text
Dual-Agent
generator–reflector architecture ensuring grading accuracy and auditability
Full Loop
trial attempts and rapid feedback enabling iterative student revision
Bedrock + S3
cloud-native foundation with granular quota governance and cost control
Ready to Transform Educational Assessment with Production Agentic AI?
Learn how Cogninest AI designs, builds, and deploys rubric-grounded agentic workflows, document intelligence, and multi-model evaluation systems on AWS.
