Safety Help That Fits What You Already Built
We build the layer that sits between your app and the AI. We use free, open tools. It runs on your computers. You own it. Then we write it all down so a buyer or a checker can see it.
Safety comes too late
Most teams wait. Then a big buyer sends a long list of safety questions. By then the AI is already live, and every little change turns into a big job.
Telling the AI to be good does not work
You can tell the AI "never share private info." That is a wish, not a rule. You cannot prove it to a checker. And it quietly fails when someone tries hard to break it.
You have no proof
A buyer asks what your AI told customer #8942 last March. Most apps never saved that, so you cannot answer.
We Add The Safety Filter
We add a scanner around your AI. We use well known free tools. Two we like are LLM Guard and Microsoft Presidio. We pick the one that fits your setup. It runs on your computers, and we are never in the middle.
- ✓Works with OpenAI, Anthropic, Azure, Bedrock, and AI you run yourself.
- ✓At first it only watches. We test it on your real traffic before it blocks a thing.
- ✓If it ever breaks, you pick what happens: let words through, or stop them.
- ✓None of your data leaves your walls. Not even to us.
- ✓The tools are free and open. If you never call us again, it keeps working.
# LLM Guard — a free tool, runs on your computers from llm_guard import scan_output from llm_guard.output_scanners import Anonymize, Sensitive scanners = [ Anonymize(vault), Sensitive(entity_types=["US_SSN", "CREDIT_CARD"]), ] sanitized, valid, score = scan_output( scanners, prompt, model_response )
It runs on your computers. Nobody sits in the middle, not even us. The tool itself is free. Setting it up right is the hard part, and that is our job.
The Filters We Can Turn On
Each one below is something the free tools can already do. We work out which ones you need, set them up for your traffic, and hand you a guide.
Hide Private Info
Names, ID numbers, card numbers, medical numbers, home addresses. We spot them and cover them up.
Stop Secret Keys Leaking
It catches passwords and keys in code the AI writes, before anyone can copy them.
Catch Unfair Choices
It flags when the AI scores people using things like age or race. Then it writes down the fix.
Block Sneaky Orders
Bad guys hide orders inside files and messages to trick your AI. We check for that first.
Mark What The AI Made
We add a hidden tag to work the AI made. The EU AI Act asks for this.
Keep A Record
Every time the filter stops something, we write it down: which rule, what we did, which AI, and when. It goes into the system you already use.
Matching You To The Rules
We set up the records so every catch is tagged to the rule it proves. Your proof piles up day by day. You are not stuck building it by hand the week before a check.
| The Rule Book | What We Check And Build | The Proof You Get |
|---|---|---|
| EU AI Act | We tell people when AI is used. We mark what the AI made. We keep records. We show a human was watching. | A record of every use, plus the history of your rules |
| GDPR | We share as little personal info as we can. We use it only for the right reason. We keep a list of what the AI touches. | A log of what we hid, and why you were allowed to use it |
| HIPAA | We hide the 18 kinds of health info. We share only what is needed. We save the reason someone looked. | A sharing log you can hand to your health partners |
| SOC 2 | Keeping what the AI says private and correct | Proof you can send straight to your checker |
| ISO/IEC 42001 | Day to day controls for running AI safely | Your rule book history and a list of problems |
We build the filters, write the papers, and save the proof. We are not lawyers, so this is not legal advice. We also cannot stamp you as approved. Your lawyer says which rules apply to you. Your checker says if you passed.
If Your AI Can Act On Its Own
An AI that only writes text is one problem. An AI that can spend money, send email or change a record is a different one. Most teams cannot say out loud what their agents are allowed to do, or how to stop one.
We go through them one at a time and write it down: every action, every key with no expiry, every missing limit, and every time one agent hands its power to another. Then we give you the fix list.
See The Agent Audit →- ✓ Every action your agents can take without a person
- ✓ Keys that never expire, and limits nobody set
- ✓ What a child agent inherits from the one that started it
- ✓ How to switch one agent off without a redeploy
- ✓ Mapped to the EU AI Act, SOC 2 and ISO 42001
Our Jobs And Prices
We agree the work and the price in writing before anything starts. No hourly bills. No monthly fee. And you own everything we make.
Help With A Buyer’s Form
3–5 days · A buyer sent you a long safety form and the deal is stuck
- ✓ The form filled in and ready to send
- ✓ A private note on every weak answer
- ✓ A list of fixes to make them strong
Full AI Safety Check
2–3 weeks · The whole picture, matched to the rules you must follow
- ✓ A list of every AI tool you run (yours to keep)
- ✓ How risky each one is
- ✓ What is missing for each rule book
- ✓ A fix list, in order, with prices
- ✓ A 90 minute call to walk you through it
Build The Safety Filter
3–6 weeks · We price each AI tool, not your whole company
- ✓ The filter turned on and tuned
- ✓ We watch first, block later
- ✓ Records saved into your own system
- ✓ A guide, and we train your team
- ✓ 30 days of tuning after we finish
Check What Your AI Agents Can Do
Every action your agents can take on their own, every key with no expiry, every missing spend limit. With a fix list.
Get Ready For The EU AI Act
We work out your job under the law, how risky your AI is, and what you must do by when.
Your Rules, Written Down
Rules for using AI. A list of your AI tools. A risk plan. What to do when things go wrong. And a form for checking other companies.
A Safety Helper Every Month
We look at each new AI feature, answer buyer forms for you, and watch for new laws.
These are close guesses. We fix the real price in writing after we plan the job, and then it does not move. We are new, so our first clients pay less if they let us tell their story. Read more →
Watch it work on your own AI
Bring one AI tool. We will put LAVA in front of it and show you what it catches, live.