Qyverix Learn
Institutions evaluate thousands of submissions with finite faculty hours and inconsistent standards. The interesting constraint was not automating the grading — it was making sure the automation never took the decision away.
Challenge
Grading video presentations and written work at scale is slow and inconsistent. Automating it naively produces a second problem worse than the first — faculty losing authority over their own assessment.
Approach
Make AI the first pass, never the last word. Every score is reviewable, adjustable and overridable with a recorded justification, and evaluation runs off the critical path so the platform never waits on it.
Outcome
Live at learn.qyverix.in as multi-tenant SaaS with institutions provisioned, faculty grading with AI assistance, and students submitting and sitting quizzes.
The brief
What the business needed.
Qyverix Learn is a cloud-native learning management platform with an AI evaluation engine sitting inside it. Institutions set up departments, programmes, courses and subjects; faculty build weighted rubrics and set video or paragraph tasks; students join by activity code and submit. The AI performs a first-pass, rubric-aligned assessment and hands faculty a criterion-level breakdown.
Two design decisions shape everything else. The first is that AI only ever suggests. Faculty review results in a dedicated console — AI scores, transcript, quantitative metrics and confidence side by side — and can accept, adjust or fully override any result before it reaches a student, with every override recording a justification. Institutions choose whether results auto-publish or stay faculty-gated.
The second is that evaluation never blocks the platform. Submissions are accepted instantly and queued; the AI engine runs asynchronously behind them. If evaluation slows down, only evaluation is affected — students still submit, faculty still teach, quizzes still run. Jobs use at-least-once delivery with idempotent handlers so retries cannot duplicate results, failures retry with exponential backoff, and anything exhausting its retries lands in a dead-letter queue with alerting rather than disappearing.
What we built
Shipped scope.
Three role workspaces
Institution admins get org structure, people management, analytics and policy. Faculty get tasks, rubrics, the AI review console and exports. Students get join-by-code, submission and feedback.
Faculty review console
AI scores, transcript, metrics and confidence side by side, with accept, adjust or override — and a recorded justification on every override.
Weighted rubrics
Criteria with weights, reusable templates and per-criterion AI hints, so the engine is scoring against what the faculty actually asked for.
Async evaluation queue
Submissions are accepted immediately and evaluated off the critical path, with automatic retries, idempotent handlers and a dead-letter queue so no job is silently lost.
Enforced tenant isolation
Every tenant-scoped query is scoped at the data layer rather than by developer discipline, extended to API guards, file storage, cache keys and background jobs — with a cross-tenant isolation suite running in CI.
Audited by design
Aligned to OWASP ASVS Level 2, TLS 1.2+, JWT access tokens with rotating refresh tokens, and an immutable audit log capturing actor, action, entity, diff, IP and timestamp on every mutating action.
Build details
How it was made.
- Stack
- Multi-tenant SaaS; FFmpeg, OpenAI Whisper, OpenCV, MediaPipe, Transformers / PyTorch
- Scope
- Product, architecture, AI evaluation engine, build, operations
- Sectors
- Education, higher education, training institutions
- Coverage
- Institution-wide, multi-tenant
- Status
- Live — P0 to P3 shipped, P4 in progress — learn.qyverix.in ↗
FAQ
Questions this raises.
Does the AI decide a student's grade?
No. The AI performs a first-pass assessment against the faculty's rubric and presents it for review. Faculty can accept, adjust or fully override any result, every override records a justification, and the institution decides whether results auto-publish or stay faculty-gated.
How fast does an evaluation come back?
The engine is designed for a paragraph submission to evaluate in under 60 seconds and a three-minute video in under eight minutes on a single CPU worker. These are design targets rather than measured averages, and turnaround tightens as GPU capacity is added.
More work