AI Email Triage: Summer Rising & Special Education Office Inboxes
NDA2026 · NYC Department of Education (Summer Rising summer program + Special Education Office)
Serverless AI system that classifies inbound parent email and drafts grounded replies for human review.
Architecture and outcomes are shareable at this level of detail; internal prompts, code, and proprietary implementation details are restricted under NDA.
A serverless AI email triage system on AWS Lambda and Amazon Bedrock (Claude), with CDK infrastructure as code. It classifies inbound parent email into 17 categories and drafts replies for human review via Microsoft Graph, grounded on Bedrock Knowledge Bases, live since April 2026 across the Summer Rising summer-program inbox and the Special Education Office inbox.
The problem
Two high volume parent facing inboxes generated more routine questions than staff could triage quickly by hand, especially during peak enrollment and program season, while sensitive Special Education correspondence needed careful handling of student and family data.
System design
Pipeline architecture
7-layer hallucination-prevention stack
Process & prototyping
- Analyzed inbound volume to find the dominant use case clusters, technical difficulties and program application questions, and built targeted RAG pipelines grounded in curated knowledge bases for those.
- Reserved the remaining scheduling and logistics inquiries, about 20% of volume, for staff handling.
- Built multimodal handling for reports that arrive with no email body, just a vague subject line and an image or video attachment, using vision and OCR to classify and route cases even when the message text is empty.
- Added PII detection and masking for the Special Education Office inbox to protect sensitive student and family data, with FERPA aligned handling.
- Designed a 7 layer hallucination prevention system (citation enforcement, similarity thresholds, regex fact grounding, Bedrock Guardrails) to keep drafts grounded in retrieved context before human review.
- SQS with a dead letter queue and EventBridge scheduling handle the serverless pipeline, deployed as ARM64 Docker images on Lambda via CDK.
Outcome & impact
Processed 13,500+ real emails across both inboxes over six months at 100% pipeline success and 94.95% auto-resolution, averaging 12.1s end-to-end latency (p95 18.3s) and 0.92 average classification confidence, across 23 detected languages.
Role & collaboration
Team
Built as part of Kshitij's ongoing NYC DOE engagement via Athreya Inc.