How We Approach AI
What runs in production today, and the architecture we plan for Claude-powered features.
Patriaworks is building AI features for its own products, starting with document summarization and information extraction. This page separates two states on purpose: the engineering that already runs, and the AI work that is still planned.
Nothing here claims a live AI product. The Claude integration described below is a build plan, and this page gets updated when the first endpoint ships.
In Production Today
- Zod schema validation on the client and on the server for every contact submission.
- A rate-limited, Turnstile-protected contact API, with structured logs that carry no credentials or message contents.
- Offline-first mobile capture with background sync, shipped in Android field apps.
- Face detection for attendance in CMP Tracker Mobile, released on Google Play.
- No large language model runs in any Patriaworks product today. That is the honest starting point.
Planned: Claude Integration
- Server-side calls to the Anthropic API from a Next.js route handler. The API key lives in an environment variable and is never bundled for the browser.
- Request validation with Zod, payload size limits, timeouts, retries with backoff, and a per-user rate limit on every AI endpoint.
- Structured outputs and tool use, so the model returns typed fields that a schema checks before anything is stored.
- Retrieval over the customer's own documents, with no document contents written to logs and no customer data used for model training.
- A review screen for consequential values, because a summary that gets a number wrong can cost more than the typing it saved.
- Token budgets and per-feature cost monitoring before launch, plus an evaluation set we run whenever a prompt changes.
Security, Privacy, and Reliability
- We process only the fields a feature needs, and our logs stay free of document contents and secrets.
- SMTP, Turnstile, and future Anthropic credentials stay in environment variables on the server, never in the browser bundle.
- A person reviews any output that feeds a real decision.
- Rate limits, timeouts, and API errors return a clear message and a retry path instead of a blank screen.
Status: not implemented yet. The environment variables, tests, and deployment steps for the first endpoint are documented in the repository before any AI endpoint goes live.
Have documents or workflows worth automating?
Send a brief describing the documents and the fields you need extracted. We will tell you plainly whether a language model is the right tool, and what it would cost to run.
Patriaworks is an independent company. Mentioning Claude and the Anthropic API means we plan to use Anthropic's public API. It is not a partnership, an affiliation, or an endorsement.
Start a Project
Tell us what the system has to do. You get a written scope and an estimate back. No discovery fee, no retainer required to talk.