Inspiration

Millions of learners write descriptive answers every day for examinations, university assessments, and professional certifications.

Modern AI can evaluate these answers in seconds. However, evaluation alone rarely changes how students learn. Most learners receive long lists of weaknesses without knowing what to improve first or how to improve it.

We wanted to answer a different question:

What if AI could become a learning coach instead of just an evaluator?

That idea became Saarthi Learning Loop.

What it does

Saarthi Learning Loop transforms structured AI evaluation into focused practice and measurable improvement.

Instead of asking learners to rewrite an entire answer, Saarthi:

Evaluates the complete answer.
Identifies the highest-impact weakness.
Converts that weakness into a focused learning objective.
Lets the learner practice only that skill.
Evaluates the revised section.
Applies the improvement back into the complete answer.
Visualizes the complete learning journey from Version 1 to Version 2.

Instead of simply telling learners what was wrong, Saarthi demonstrates exactly how they became better writers.

Why this matters

Traditional AI grading ends when the evaluation is complete.

Saarthi begins there.

The goal is not merely to generate feedback but to create a deliberate practice loop that learners can repeat every day.

This approach encourages measurable skill development instead of one-time assessment.

Although this demonstration uses UPSC-style descriptive answers, the learning loop can be extended to essays, legal writing, university examinations, language learning, and any domain where structured writing and revision are important.

How we built it

Saarthi Learning Loop was implemented as an OpenAI Build Week extension on top of our existing MainsCraft platform.

The Build Week implementation introduced:

dedicated Build Week experience
highest-leverage learning diagnosis
focused micro-practice
section-level evaluation
one-section Version 1 to Version 2 application
Improvement Journey visualization
Build Week documentation
deterministic validator
submission assets

The existing platform already provided authenticated workflows, answer persistence, and GPT-5.6-powered structured evaluation through a private Eval Engine.

GPT-5.6

GPT-5.6 powers structured educational evaluation through our private Eval Engine.

It provides:

structured answer evaluation
educational reasoning
answer evolution
feedback generation

For the Build Week demonstration, the section-level practice flow intentionally uses deterministic demo logic to ensure a reliable, repeatable demonstration without uncontrolled provider calls.

How Codex helped

Codex played a central role throughout the implementation.

It was used to:

inspect the existing platform
understand the architecture
design the Build Week extension
implement new React components
integrate routing
create validators
prepare documentation
generate submission assets
support testing and quality assurance

Challenges

The biggest challenge was balancing educational value with a short demonstration.

Instead of adding more features, we focused on one complete learning loop that clearly demonstrates measurable improvement.

Another challenge was ensuring the demo remained deterministic while accurately representing the role of GPT-5.6 within the production architecture.

What's next

Our long-term vision is to extend Saarthi Learning Loop beyond civil service preparation into a general AI-powered deliberate practice platform for any domain that requires high-quality written reasoning.

Future work includes richer adaptive learning plans, personalized skill progression, broader educational domains, and deeper integration with our private evaluation engine.