MEET YOUR FIRST BOUNDED AI
A little AI.
A clear boundary.
It can sort a request and carry out a small task.
You decide what it’s allowed to do.
Try it in a minute. No account needed.
01 / GIVE IT SOMETHING TO UNDERSTAND
Where should this request go?
Try an example. This small AI suggests a support team.
02 / TURN A REQUEST INTO AN ACTION
Give it a small job.
Zyvora can use tools, too. Start with one you can check yourself: counting words.
03 / YOU SET THE BOUNDARY
It asks before it acts.
This time, the word-counting tool isn’t authorized. Start the task and see where it stops.
That’s Zyvora.
Understand a request. Use a tool. Respect your decision.
A small, real prototype — not a general-purpose AI assistant. It runs locally on the server and never trains on your input.
Real model. Real execution. You stay in control.
Curious? Open the workbench
Plan. Authorize. Execute.
Unchecking authorization demonstrates the approval gate. A decision expires after 5 minutes. Rejection, timeout or execution failure ends the task. No retries or restart recovery.
Every task has an ending.
Keep only what helps.
Task inputs live in server memory for up to 10 minutes. Only route, confidence and a hashed session key enter MySQL; request text is never stored there. History expires logically after 24 hours and is pruned on new inference.
Inspect the real system.
JVM metrics are shared across this application's process. Model scores are uncalibrated toy probabilities, not accuracy guarantees.
The boundary is part of the product.
Model & research notes
Multinomial Naive Bayes with Laplace smoothing; 24 hand-authored synthetic training examples, four equally represented labels. Independent phrase fixtures verify routing and abstention. No user input is used for training. Low vocabulary coverage or confidence below 0.65 routes to review. This is a bounded rule planner with statistical inference, not a general autonomous agent or an academic novelty claim.
Runtime internet retrieval, paid models, physical devices, arbitrary code, self-upgrades and multimodal inference are disabled. Data export supports JSON, CSV and password-encrypted JSON; reports and other output formats remain unverified adapters.
Architecture follows the existing IAFM Controller / Model / DatabaseService structure. Research sources: Tomcat deployment · Stanford Naive Bayes.