Prompt-Injection Defences for LLM Features in Rails
A practical guide to defending Rails apps against prompt injection: limiting what the model can do, treating its output as untrusted, gating side effects and testing for attacks.
Software that works, by the date you need it.
Event software, connected hardware and web products, built on a fixed scope by a senior engineer. AI speeds up the work. Fifteen years of experience keeps it rock solid.
I'm Wagner Matos, based in Dorset and working remotely across the UK, US and Europe. Past clients include Johnson & Johnson, GSK, Citizens Advice, Babbel and Carwow.
-
Building for a congress stand?
Touchscreens, photo experiences and stand analytics, on the floor when the doors open.
-
The cloud, updates and remote access for devices on networks you do not control.
How it works
-
Step 1
A short call
Thirty minutes, free. Enough to tell whether I can help.
-
Step 2
For bigger builds, after the call: one paid week that turns your idea into a spec, an architecture and a fixed price.
-
Step 3
Build, launch and support
Working software at every milestone, then deployed, documented and supported.
Not a build? I also take contract engagements and fractional support. More on the services page.
What clients say
When we hand Wagner a build with a hard deadline - a congress opening, a live event - it ships. Complete, working, on the floor when the doors open.
Wagner was enjoyable to work with. He was able to quickly get up to speed, immediately wrapping his brain around problems and shipping clear code to fix them.
Wagner handled our blog migration end to end including migrating all content, URLs and redirects for SEO, and a custom Cloudflare Worker. All documented and deployed with zero drama.
Cost Controls for LLM Features in Rails: Per-Tenant Token Budgets, Caching and Model Routing
How to keep the API bill for LLM features in a multi-tenant Rails app predictable, using per-tenant token budgets, response and prompt caching, and routing each request to the cheapest model that can handle it.
Evaluating LLM Features in Rails: Regression Tests for Prompts with Minitest and Fixtures
AI features in Rails can quietly get worse when you change a prompt or a model. This guide shows how to catch that with Minitest, fixtures, snapshots and a small eval suite.
Building an MCP Server Inside Your Rails App
A hands-on guide to exposing your Rails app's data and actions to AI assistants with the official Ruby MCP SDK: tools, authentication, Streamable HTTP and testing.
Stay up to date
Get notified when I publish something new, and unsubscribe at any time.
Recent Work
Citizens Advice
Citizens Advice helps millions of people a year with debt, benefits, housing, consumer and energy problems. Casebook is the Rails application its advisers use to do that work: every client, every case, every appointment, across the national network. I joined the engineering team as one of several senior developers and worked on the core of it.
Johnson & Johnson
During the ECP Congress 2025, Johnson & Johnson needed a way to bring their precision medicine story to life on the exhibition floor. Healthcare professionals would be passing through quickly, the experience had to attract, educate, and track engagement, all without depending on venue Wi-Fi or cloud infrastructure.
Confidential (healthcare)
I built a serverless webhook service that delivers real-time conversation events to each NHS trust's own systems, handling encryption, authentication, concurrent delivery, and automatic retries.
Johnson & Johnson
For Johnson & Johnson Medical Affairs at a major international medical congress in 2025, I built an interactive presentation kiosk that works with no network at all, and still hands the team a full picture of what visitors looked at, for how long, and how far they got. The fourth build for the same client team, across three congresses.