EdTech multimodal grading
OCR, rubrics, and a custom grading API on one platform.
Multimodal pipelines for worksheet capture, rubric enforcement, and explainable scores — shipped as a governed `/v1` route.
A curriculum vendor needed structured scoring from handwritten and PDF submissions. Perception pipelines normalize scans; a custom dex28 route applies the rubric and returns machine-readable results plus human-readable rationale.
The same platform powers sandbox trials and contract-grade throughput.
Ingress → perception (OCR/layout) → rubric service → LLM adjudication → durable result store. Each hop is observable; failures surface as partial grades with retry hints rather than silent drops.
Embeddings index exemplar answers for consistency checks across graders.
FERPA-aligned data handling with explicit retention windows; keys scoped to school tenants. Audit logs record who authored rubric versions and who approved deploys.
No student content trains shared models without a separate, signed data plan.
Time-to-first accurate rubric dropped from weeks to days; product could A/B models without re-plumbing. Support load fell because teachers saw the same traces engineers used — fewer “black box” escalations.
Net: predictable cost per thousand submissions with a clear upgrade path as models improve.