About Dwi
A small human layer, developed through practice.
Dwi is an independent open-source project by Trần Thiện Học. It turns repeated lessons from working with AI agents into modules that people can read, try and remove.
What Dwi is
Dwi sits around an existing agent workflow as a set of inspectable instructions.
It helps make the intended outcome, permitted scope, ownership, resource boundary and supporting evidence easier to see.
It is not another model, a hidden runtime or a replacement for native permissions.
The perspective behind the project
Different disciplines taught different parts of the same lesson.
Technology taught dependencies.
Design taught proportion.
Communication taught how meaning can be lost between layers.
Product work taught that a capable system can still leave people unsure what to do.
Leadership taught that unclear ownership becomes more expensive as work scales.
Working with AI brought these lessons together.
Dwi began with a practical question: can the work carry more of its own memory, boundaries and evidence so that the person does not have to carry everything alone?
A background across systems and stories
For more than a decade, Thiện Học has worked across strategy, communication, digital experience, product collaboration and team leadership.
The work has taken place in fast-changing environments where clarity, trust and coordination could not be assumed.
Rather than presenting that history as a list of achievements, Dwi uses what the work taught: structure before spectacle, meaning before amplification and alignment before acceleration.
The question Dwi explores
How can AI systems move faster without making the person responsible for the result carry more invisible work?
Dwi approaches this through five connected concerns:
- intention;
- proportion;
- resources;
- coordination;
- evidence.
The goal is not to remove people from the process. It is to return human attention to the moments where judgment matters.
Practice-led research
Observe.Separate.Build.Test.Revise.
Dwi begins with recurring pressure found in real work, not with the assumption that every workflow should operate in the same way.
- Observe a recurring problem.
- Separate the human burden from the model’s visible behavior.
- Define the smallest operating change that may help.
- Turn it into an inspectable module.
- Test it on a bounded, reversible task.
- Preserve failures and unknowns.
- Revise when evidence changes.
- Publish the limitation beside the benefit.
If an artifact cannot be inspected or removed, it does not fit the project’s method.
Experience creates questions.It does not replace evidence.
The author reports more than 20 billion tokens of AI use across coding, research, product work, writing and agent workflows.
This describes the breadth of experience behind Dwi. It does not prove universal savings, safety, compatibility or quality.
Every product claim still needs a defined case, source, date, measurement boundary, limitation and evidence state.
Project principles
These principles keep the project small enough to inspect, test and remove.
- Human authority stays explicit.
- Small tasks should remain small.
- Resources should be visible.
- Advice is not permission.
- One changing scope has one writer.
- Claims carry evidence labels.
- More process must justify its cost.
- Removal should remain possible.
- Failures and unknowns stay visible.
- Respect for the user matters more than performance theater.
Dwi is an independent open research preview. Actual behavior depends on the agent tool, model, project instructions, permissions and task.
Dwi by thienhoc
Careful critique is welcome.
Dwi should become more useful through testing, not more convincing through repetition.