REF / AUTOMATION

California EdTech - AI Grading and Parent Reporting Automation

Built AI-graded assignment workflows and automated parent reports for a California-based K-12 edtech platform - saved teachers 11h/week each.

RoleAutomation Lead
Year2025
Outcome+11h/week per teacher
DomainAutomation
00
STACK

Tech used.

Claude Sonnet 4.6Postgresn8nResendStripeWebhooks

The Problem

A California-based K-12 supplemental learning platform serving ~28K students was promising teachers personalised insights into student progress. but the back end was running on goodwill. Teachers were grading written responses and short-answer questions manually, often during evenings and weekends. Parents wanted weekly progress updates; teachers either didn't have time to write them, or the platform delivered generic auto-generated templates parents quickly stopped reading.

The product was good. The teacher experience was burning the staff out.

What I Built

1. AI grading for written responses. A Claude-powered grader processes every short-answer and written-response submission against the teacher's rubric (ingested per-assignment). Outputs include: a numerical score, specific praise (1-2 sentences citing the strongest element of the answer), specific actionable feedback (1-2 sentences on what would improve it). Teachers review and approve/edit; their hands-on time per assignment dropped from 4-5 minutes to 30 seconds.

2. Personalised parent reports. A weekly automated email per family summarises the student's week: topics studied, mastery progress, specific moments of strength, areas teacher recommends practice, suggested at-home activities. Generated by Claude from the student's actual work, not templates. Open rates went from 31% (template era) to 79%.

3. Teacher dashboard. Daily AI-curated insights for each teacher: "5 students need re-engagement", "3 students mastering new concepts faster than expected", "Class struggling with topic X. here are two suggested re-teaching approaches."

4. Math reasoning verification. For math problems requiring shown work, the AI checks not just the final answer but the reasoning steps. surfacing students who got the right answer through wrong logic (a category teachers consistently said they wished they had time to catch).

Submission to Parent EmailAI proposes, teacher approves
  1. 01
    TriggerStudent submits assignment

    Written response, short answer, or shown-work math problem lands in the platform.

  2. 02
    StepAI grades against rubric

    Claude scores against the teacher's rubric and produces 1-2 sentences each of specific praise and actionable feedback.

    4-5 min → 30s
  3. 03
    DecisionTeacher reviews & approves

    Teacher edits or accepts. Nothing reaches the student until a human signs off.

  4. 04
    StepGrade + feedback released

    Student sees results within 24h instead of 3 days. Daily teacher dashboard surfaces struggling-class patterns.

  5. 05
    OutputWeekly parent email

    Per-family summary built from the student's actual work: topics, mastery, strengths, recommended at-home practice.

    31% → 79% open rate
+11h
Teachers' weekly time recovered
79%
Parent report open rate
from 31%
+24%
Student weekly engagement
post-rollout
8.4 → 9.6
Teacher NPS
out of 10
Teacher Workload
MetricBeforeAfterΔ
Hours/week on grading9h1h−8h
Hours/week on parent comms4h30m−3.5h
Avg feedback turnaround on assignment3 days<24h−66%
Personalised insight quality (parent rated)2.8/54.4/5+1.6
Teacher monthly burnout flag rate27%8%−19 pts

Outcome

Teacher retention. the platform's biggest pre-existing risk. improved measurably. The internal teacher NPS climbed from 8.4 to 9.6. Student engagement went up 24% (faster feedback loops drive learning loops). Parents started referring other families because they could finally tell, in plain English, what their child was actually learning each week.