AI in production at a $5M nonprofit

In 2023, an internal theft incident involving residents' mail exposed a control gap. I immediately reduced the risk of theft with policy and culture changes, but we still had both the technical and labor-intensive challenge of sorting and moving 250+ pieces of mail through the facility daily.

Front-desk staff photograph each letter or shipping label. A vision agent reads it, resolves the recipient across three separate Dynamics 365 environments (shelter, treatment, sober living, with recency-based routing when a name matches in more than one), notifies the person directly, and sends per-program digests to staff. Everything runs in the org's own workspace with vaulted credentials. The package-tracker half alone carries 232 automated tests and runs as an hourly weekday scheduled agent.

The organization receives 80% of donations as paper mail. The processing of the documents became a bottleneck that would slow down fundraising and compliance. It took hundreds of staff hours a year in manual data entry, and our donor stewardship efforts depend on accuracy.

An AI pipeline on Azure AI Foundry that automates acceptance, recording, and acknowledgment. Staff take a picture of the donation and the AI agent automatically processes it. I started with consultants on the proof-of-concept, then brought iteration in-house. Accuracy is checked at several layers: cross-model OCR verification, inverted confidence scoring to surface hard-to-read image regions, and a custom DeepEval benchmark suite built for this exact workflow, re-run against each candidate model (four generations so far) before anything is changed.

The frontier, or jagged edge, of AI capabilities is not user-friendly. I needed to solve how to disseminate the latest knowledge and tools for using these systems to my team.

I packaged my own working toolkit into a distributable plugin and published it organization-wide: 11 skills and 10 commands covering writing, structured analysis, spreadsheet building, policy drafting, document QA, and web research. It ships in two form factors so adoption reaches both the technical and non-technical staff. Two highlights: a writing-quality skill built on 67 documented patterns of AI-sounding prose, so staff output meets our writing standards instead of being AI-slop; and a competitive-pay agent that sweeps local job postings against our 16 tracked roles and syncs matches into the SharePoint list HR uses to set floor pay, turning an annual manual comp survey into an automated system.

Clinicians spend too much time writing therapy notes, and most AI shortcuts on the market move protected health information out of our control, do not integrate with our existing systems, or are too expensive.

A documentation pipeline that turns clinical session transcripts into draft progress notes entirely inside a signed-BAA environment to ensure we are compliant with HIPAA. Staff record sessions and get accurate and compliant clinical notes sent to them quickly. Client and staff consent is managed, transcripts auto-delete, and output files as a factual progress note. I built it for the clinical group practice where I serve as practice administrator, and a pilot inside our own residential treatment program has been green-lit.

Commercial e-signature quotes came in at $7,200–$10,800/yr for our ~2,250 annual documents, with most solutions being overkill for our needs.

A compliant and useful DocuSign-like app for staff plus a managed signature API for our client documentation automations, with an AI agent to keep it running. A self-hosted platform handles the ~250 human-initiated HR and volunteer documents while a hosted API handles the ~2,000 machine-initiated client documents. It runs on the org's self-hosted application server, which I hardened myself: 2FA on every admin account, vaulted secrets, and encrypted offsite backups. An AI agent provides the sys admin work to keep it working.