Hours of paperwork, turned into minutes of review.

Context

SheltrLink’s caseworkers were spending hours per client just writing housing plans — synthesizing intake-form answers into structured narratives that admins could review and approve. The intake-to-plan pipeline was the bottleneck for how many clients the team could serve in a week.

Approach

Built a structured intake form that captures the client’s situation in a typed, validated shape (using Prisma + MongoDB for flexibility on the document side). When intake is submitted, the API hands the structured data to OpenAI with a tuned system prompt that produces a draft housing plan tailored to the client’s circumstances. The plan goes to admins as a draft they can edit and approve — AI is the first draft, never the final word. There’s also an AI-suggestion layer for referral matching, surfacing relevant partner organizations from a knowledge base of past placements. Real-time updates flow over Socket.io so caseworkers see status changes without refreshing. Approved plans export to PDF via Puppeteer + PDFKit. JWT-based authentication for caseworker access; role-scoped admin review.

Outcome

Caseworkers now spend their time reviewing and refining plans instead of writing from scratch. Admin throughput went up. Quality stayed consistent because every plan starts from the same well-structured prompt.

What it does

• 01AI-generated first-draft housing plans from structured intake forms (OpenAI integration with tuned system prompts and structured output parsing)
• 02Real-time collaboration via Socket.io across caseworker and admin views
• 03Caseworkers shifted from authoring plans to reviewing them — with quality held steady by a single, well-tuned prompt