How did I build SudoChat?

Author: Mustafa Siddiqui · Part: 3 of 3 · Recruiter question: How was SudoChat built?

Evidence rule: This is my first-person authored source material. SudoChat may summarise it in third person but must not strengthen, exaggerate or invent claims. Limitations and Do not claim sections are hard factual boundaries.

## Simplified SudoChat architecture
At a high level, SudoChat can be described as:

**User**

↓

**SudoChat frontend**

↓

**Copilot Studio conversational layer**

↓

**Instructions and responsible AI boundaries**

↓

**Knowledge retrieval**

↓

**Curated evidence about me**

↓

**LLM generates a grounded response**

↓

**Response returned to the user**

Alongside this is an evaluation and governance layer:

**limitations + do not claim rules + fallback behaviour + testing**

The important point is that the LLM is only one component.

The quality of the system depends on the knowledge, retrieval, instructions, user experience, evaluation and boundaries surrounding it.

## Why I built it this way
I could have built a simpler demonstration that sent prompts directly to a language model.

I deliberately chose not to.

The project is intended to demonstrate several of the concepts I believe are important for enterprise AI:

* grounding AI in controlled information
* separating verified evidence from model knowledge
* reducing hallucination
* defining what a system should not claim
* designing safe fallback behaviour
* making knowledge maintainable
* testing AI behaviour
* making the system understandable to users
* thinking about the entire system rather than only the model

That makes SudoChat both a portfolio project and an engineering experiment.

## Relevance to the Federal Courts
SudoChat is relevant to the Court AI Technologist role because the underlying problem resembles many enterprise AI use cases.

A Court assistant would also need to answer questions using controlled information rather than simply trusting an LLM's general knowledge.

The actual Court implementation would be substantially more complex, but the same principles could apply:

**Authorised Court user**

↓

**authenticated AI interface**

↓

**controlled knowledge sources**

↓

**retrieval**

↓

**grounded response**

↓

**sources and evidence**

↓

**safe refusal when information is insufficient**

The Court version would additionally need controls around areas such as:

* identity
* permissions
* information classification
* privacy
* audit logging
* records management
* security
* source authority
* human review

SudoChat does not claim to solve all of those problems.

It demonstrates my ability to think about them and convert responsible AI principles into a working proof of concept.

## Limitations or gaps
SudoChat is a proof of concept.

It should not be represented as a production enterprise system.

It should not be represented as a Federal Court system or as using Federal Court internal information.

The initial knowledge base is manually curated and structured.

That is intentional for the proof of concept because I want strong control over what evidence the assistant initially receives.

The project is still being expanded and evaluated.

Its direct Copilot Studio implementation experience is recent.

SudoChat should therefore be presented as evidence of practical learning, architecture thinking and responsible AI experimentation rather than evidence that I have already operated a large scale Copilot Studio environment.

## Useful links
SudoChat project page

SudoChat GitHub repository

My portfolio:
https://mustafa-siddiqui.com/

GitHub:
https://github.com/sudoqui

Federal Court AI Technologist application evidence base

Responsible AI research

Privacy preserving government chatbot research

## Do not claim
Do not claim SudoChat is a Federal Court application.

Do not claim SudoChat has access to Federal Court systems or information.

Do not claim it is a production enterprise deployment.

Do not claim RAG eliminates hallucination.

Do not claim the assistant can verify every answer with certainty.

Do not claim SudoChat was fine tuned on my personal information.

Do not claim the LLM itself contains the SudoChat knowledge base.

Do not claim the current proof of concept implements every security control that would be required in a real Court environment.

Do not claim SudoChat represents the Federal Courts' existing AI architecture.

Do not claim the Engine Room represents physical infrastructure.

Do not claim the manually curated knowledge base would necessarily be the final approach for an enterprise Court implementation.

The accurate representation is that SudoChat was built as a **bounded Copilot Studio proof of concept combining a custom frontend, structured evidence, retrieval and grounding, explicit claim restrictions, uncertainty behaviour and responsible AI principles to demonstrate how an assistant can be useful without being encouraged to invent information.**

← Previous part · Knowledge Base