ZynoviqBuilt your own agent? Test it like a professional would.
Many teams build their own finance or tax agents — on a general-purpose AI model, with their own prompts, tools and documents. They are fast to build and hard to check. We test them the same way we test any system: realistic cases, certified answers, every failure explained.
What we can test
- Tax and compliance assistants that answer questions in plain language
- Agents that read documents and propose a treatment or posting
- Multi-step agents that call tools, look things up and then decide
- Retrieval-based assistants grounded in your own policies or manuals
Connecting your agent
| Method | How it works |
|---|---|
| API | We send each case to your test endpoint and receive a structured answer |
| Agent protocols | Where your agent supports standard agent-to-agent or tool protocols, we connect through them (such as MCP and the Agent2Agent protocol) |
| File exchange | We send a batch of cases; you return a file of answers |
The answer format
For exact, fair marking, your agent returns fixed fields alongside any explanation — for example:
{
"case_id": "GST-ITC-00417-V07",
"itc_eligible": "yes",
"eligible_amount_inr": 10000,
"reverse_charge": "no",
"legal_basis": ["CGST s.17(5)(b) proviso"],
"explanation": "free text"
}If your agent only answers in plain language, we can mark by professional review instead, at a different price.
Useful extras for builders
- Failure patterns tied to causes you can act on — missing exceptions, outdated rules, wrong facts.
- Consistency checks — the same case asked in different words, to find unstable answers.
- Confidence honesty — whether your agent says "I'm not sure" when it should.
- Improvement data in your format when you want to fix what failed. Improvement data.
What we never do
We never run your model ourselves, see its weights or training data, or use your agent's outputs to train anything. What we do and don't do.
Testing custom AI agents — FAQ
Which AI models can your tests work with?+
Any. We test your agent's answers, not the model behind it.
Do you check whether our agent answers consistently?+
Yes. The same legal point is asked in different ways to find unstable answers.
Do you ever run our model or see its internals?+
No. We never run your model, see its weights or training data, or use its outputs to train anything.