AI Email Triage: production system + live demo
NDA2026 · NYC Department of Education (Summer Rising + Special Education Office)
A serverless AI email triage system I shipped for NYC DOE, plus a public, synthetic demo of the same pattern that anyone can run live.
The production system's architecture and outcomes are shareable at this level; internal prompts, code, and proprietary details are restricted under NDA. The live demo below is a separate, fully synthetic rebuild of the pattern, with no real data in it.
In production I built 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 grounded replies for human review via Microsoft Graph, grounded on Bedrock Knowledge Bases, live since April 2026 across the Summer Rising and Special Education Office inboxes. That work is under NDA, so I also built MailRoom: a public, fully synthetic rebuild of the same pattern that anyone can run live, on open models (a local model on my own GPU with a hosted Llama fallback) and a RAG pipeline written from scratch. Every email, name, and policy in the demo is invented.
Live demoRun the whole pipeline in your browser: classify, retrieve, draft, and seven guardrails, on synthetic data. Nothing is stored or sent.Open the live demoThe 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. And because the production system is under NDA, I needed a way to show the pattern actually running without exposing anything proprietary. That is what the live demo does.
System design
Pipeline architecture
7-layer hallucination-prevention stack
Process & prototyping
In production (under NDA)
- Targeted the dominant clusters firstAnalyzed inbound volume to find where it concentrated, technical difficulties and program application questions, and built RAG pipelines grounded in curated knowledge bases for those.
- Left the tail to peopleThe remaining scheduling and logistics inquiries, about 20% of volume, stay with staff.
- Multimodal handling for empty emailsReports often arrive with no body at all, just a vague subject line and an image or video. Vision and OCR classify and route those cases anyway.
- PII masking for Special EducationDetection and masking to protect sensitive student and family data, with FERPA aligned handling.
- Seven layers against hallucinationCitation enforcement, similarity thresholds, regex fact grounding, and Bedrock Guardrails, keeping every draft tied to retrieved context before human review.
- Serverless pipelineSQS with a dead letter queue and EventBridge scheduling, deployed as ARM64 Docker images on Lambda via CDK.
In the public demo (MailRoom)
- Same pipeline, open modelsRebuilt from scratch on a local model on my own GPU with a hosted Llama fallback, over a Chroma RAG pipeline written by hand instead of a managed one, and runnable with no API key so anyone can watch every stage and guardrail run on synthetic data.
- Zero cost hosting, measured metricsAzure Container Apps on a scale to zero plan so it costs nothing when idle, reaching the home GPU over an authenticated Cloudflare tunnel, and instrumented to report its own numbers rather than invented ones.
Outcome & impact
Every number below is measured by the code itself rather than estimated.
In production (six months, both inboxes)
In the public demo (18 synthetic emails)
Role & collaboration
Team
The production system was built as part of my ongoing NYC DOE engagement via Athreya Inc.
My role
The public MailRoom demo is a solo build: pipeline design, backend, RAG, the seven layer guardrail system, and deployment.