Skip to main content
The open-source framework is live today. The handover is in early access, and we are taking on our first design partners. Talk to us about early access.
Each task moves to your own AI only after it passes a test. If quality slips, it moves back automatically, and every step is recorded for regulators.

​
How it works

  1. Watch: records how each task is done today.
  2. Test: your own AI takes an exam made from real past work it never studied. The work is split first: 70% to study, 30% kept back as the exam, never mixed. It must reach a pass mark your risk team sets in advance, such as 99%, proven on at least 400 exam cases.
  3. Hand over: after 100 practice answers and a named person’s approval, a 10% trial with every answer checked, then all of it.
  4. Roll back: if the checked answers drop below 95%, errors pile up, or anything changes, the task moves back automatically.
  5. Record: every step goes into a sealed diary with no customer text, which your regulator can check.
The numbers are the handover’s strict defaults. No regulator sets one accuracy number for AI: the EU AI Act asks for an “appropriate level of accuracy”, and UK bank supervisors ask the board to set its own thresholds. So your risk team sets the pass mark for each task in advance, and it is written in the record.

​
Where your own AI comes from

Your own AI is an open model, such as Llama, Qwen or Gemma, running on your own computers. There are two ways to get one:

Use an open model as it is

Serve it with Ollama, LM Studio, llama.cpp or vLLM. Available today.

Train one on your own past work

PraisonAI Train fine-tunes a small open model on your own NVIDIA GPU servers. Turning watched work into training data is part of early access.
Either way, the model is only a candidate. It takes over a task only after it passes the exam: real past work it never studied, scored against the answers your business already accepted. Your team starts each training run; nothing trains automatically.
Answers written by another AI provider’s model can be covered by that provider’s terms of use, which may forbid training another model on them. Having your reviewers check those answers does not remove the terms. Check your provider’s terms before you train. The safer source is answers your own people wrote or your own open models produced.
What is “your own AI”? An open model, such as Llama, Qwen or Gemma, running on your own computers (fully offline, on local model servers you host). Use it as it is, or train it on your team’s past work with PraisonAI Train on your own GPU servers (NVIDIA GPU required). Your team decides when to train; nothing trains automatically. Either way, a task moves to it only after it passes its tests.

​
Who it is for

Banks, insurers and hospitals that want AI agents but cannot send private data to outside AI services. Keep your data inside, own your AI, and get proof for your regulators.

​
What it builds on today

The handover is not part of the open-source packages. It is built on parts that are already in the open-source framework, which you can use now:

Tracing

Record what an agent did, step by step

Evaluation with run_trials

Run each test case many times and get a pass rate

Fine-tuning with praisonai-train

Train a model of your own on your own data

Local models

Run agents fully offline on your own computers

Model router

Choose which model handles each task

Local model servers

Ollama, LM Studio, llama.cpp and vLLM

​
Early access

We are taking on our first design partners. If you run a highly regulated team, such as a bank, insurer or hospital, and want to try the handover, get in touch.
No. It is in early access and is not part of pip install praisonai. Everything else on this site is open source and available today.
No. Training is optional, and your team chooses when to run it. You can also use an open model as it is.
Yes. Tracing, evaluation, fine-tuning and local models are all in the open-source framework. The cards above link to each one.
No. PraisonAI is still a general-purpose agent framework, and every feature works as before. See the Introduction.