Divyakush Punjabi
GovernAI Studio — AI governance simulator

GovernAI Studio

Interactive training simulator for governance and compliance scenarios, backed by a RAG and LLM inference path.

Also
Simulator logic
Inference
Hybrid RAG + LLM
Async
Celery workers

What it had to solve

Governance and compliance are learned by judgement, not by recall, which makes them hard to teach from a document. A simulator has to hold a scenario, respond in the voice of a specific stakeholder, and stay anchored to the actual policy rather than improvising plausible-sounding rules — while the inference doing that work takes long enough that it cannot block the interface.

The decisions that shaped it

Persona-driven interface

The front end is organised around personas rather than around forms: the learner is talking to a stakeholder in a scenario, so the interface has to sustain a conversation and hold state across turns instead of collecting a submission.

Retrieval before generation

A hybrid RAG and LLM path means the model answers against retrieved source material rather than from its weights alone. In a compliance trainer that is the difference between a tool and a liability — the responses have to trace back to the governance documents that were actually loaded.

Celery for work that outlives a request

Inference and retrieval run as Celery tasks off the Django application, so the request cycle is not held open behind a model call and the interface can show progress rather than a spinner on a blocked connection.

Extending the simulator, not just skinning it

Beyond owning the front end I extended the core simulator logic in the backend, which is what kept the interface and the scenario engine from drifting into two different models of what a scenario is.

How the pieces sit together

Client
React + TypeScript, persona-driven scenario interface
Application
Django serving the simulator and its scenario logic
Workers
Celery running retrieval and inference off the request cycle
Inference
Hybrid RAG + LLM path grounded in loaded source material

Built into the product

Scenario simulation

Governance and compliance situations played out in conversation.

Persona interface

Responses in the voice of a specific stakeholder.

Grounded responses

Retrieval-augmented generation over real policy material.

Asynchronous inference

Celery workers so long calls never block the UI.

What it is built out of

Client
React, TypeScript
Backend
Django, Celery, Python
Inference
RAG, LLM inference

What it looks like

GovernAI Studio — The simulator entry point — what the training covers, before a scenario is picked.
The simulator entry point — what the training covers, before a scenario is picked.
GovernAI Studio — Enrolment: the credential logged before the first simulation runs.
Enrolment: the credential logged before the first simulation runs.
GovernAI Studio — The scenario list — each one a dilemma, unlocked in order.
The scenario list — each one a dilemma, unlocked in order.
GovernAI Studio — Training complete — concepts mastered, partial and still to revisit.
Training complete — concepts mastered, partial and still to revisit.

What it produced

RAG
Grounded rather than free-generated answers
E2E
Interface built end to end
Async
Inference moved off the request cycle