Introducing XASRI Taksh-01

Today, we’re introducing XASRI Taksh-01, the first model from XASRI’s engineering-focused model development program.

Taksh-01 was built with a specific objective: to move beyond general-purpose AI assistance and explore what it takes to build a model that can understand and work with engineering problems, technical requirements, constraints, design decisions, and implementation workflows.

This first training phase has now been completed.

What We’re Building

Taksh-01 is being developed as an intelligence layer for XASRI’s broader engineering systems.

The long-term goal is not simply to build a model that can answer engineering questions.

We want to develop a system capable of working through an engineering problem from its initial requirements toward a validated implementation.

That includes areas such as:

Understanding engineering intent and requirements Working with technical constraints Planning engineering tasks Reasoning about design decisions Supporting calculations and technical analysis Working with research and engineering evidence Planning 3D and parametric designs Generating and working with engineering software Supporting hardware-oriented development Working with platforms such as Arduino, Raspberry Pi, and NVIDIA Jetson Connecting engineering decisions with testing and validation

Taksh-01 is one part of that larger system. The model itself will work alongside specialized engineering tools, research systems, coding systems, CAD and simulation workflows, and validation layers.

What We Learned From Phase 1

The initial benchmark results were encouraging.

They showed that Taksh-01 was learning the capabilities we were targeting during training.

But our evaluation also showed something equally important:

Benchmark performance alone is not enough to demonstrate complete engineering capability.

When we looked beyond benchmark-style tasks and considered the broader requirements of an engineering workflow, there was still a significant gap between the current model and the level of reliability we expect.

That gap is exactly what we need to work on next.

We are not treating the first training run as the finished model.

We are treating it as our first meaningful baseline.

What Comes Next

We are now preparing Phase 2 of Taksh-01 training.

The next phase will focus on improving the areas where the first version needs more capability, particularly:

Technical reasoning Improving the model’s ability to reason through engineering problems rather than simply produce plausible-looking answers.

Engineering context Improving its understanding of requirements, dependencies, constraints, and the relationship between different parts of an engineering task.

Consistency Making its outputs more reliable across different problems and longer workflows.

Constraint handling Improving how the model works when multiple technical, dimensional, manufacturing, software, or system constraints need to be considered together.

Engineering workflow understanding Moving from isolated responses toward reasoning across connected stages of an engineering problem.

The objective of Phase 2 is therefore not simply to improve benchmark numbers.

The objective is to make Taksh-01 more useful for real engineering work.

From Model to Engineering System

Our vision for XASRI is larger than the model itself.

Taksh-01 will eventually operate as part of an engineering environment where the model can work with specialized systems for research, calculations, design, CAD, simulation, coding, testing, and validation.

For example, an engineering task could eventually move through a workflow such as:

Engineering Intent → Requirements → Research → Engineering Reasoning → Calculations → Design → 3D/CAD → Simulation → Validation → Software/Hardware Implementation → Testing

The model provides intelligence across this workflow, while specialized tools perform the operations that require deterministic computation, geometry processing, simulation, code execution, or validation.

This distinction is important to us.

We don't want an AI that simply looks like an engineer.

We want to build the foundations for an AI engineering system that can reason, use tools, produce artifacts, test its work, identify failures, and improve its designs.

The Beginning of Taksh-01

Taksh-01 is not the final version of what we are building.

It is the first step.

The first training phase gave us a baseline. The evaluation gave us a clearer understanding of the model’s strengths and limitations. Phase 2 will use those findings to push the model further.

There is still a lot of work ahead.

But now we have something important that we didn't have before:

a trained engineering-focused model, a baseline to measure against, and a much clearer understanding of what needs to improve.

XASRI Taksh-01 — Phase 1 complete. Phase 2 ahead.