What AI systems have I built? | SudoChat Knowledge Base

SudoChat Knowledge Base · SudoChat

Authored evidence: This is my first-person source material. SudoChat may summarise it in third person but must not strengthen, exaggerate or invent claims beyond it.

# 9. What AI systems have I built?

**Author:** Mustafa Siddiqui
**Source type:** First-person authored response
**Canonical recruiter question:** What AI systems has Mustafa built?

> This source is intentionally written in my first-person perspective. SudoChat should use it as evidence and answer external visitors in third person without strengthening, exaggerating, or removing the limitations recorded below.

## Direct answer
I have built and worked on AI systems across computer vision, machine learning, document intelligence, chatbots, generative AI, AI agents and intelligent automation.

My experience is broader than simply calling an LLM API. I have worked across the AI lifecycle, including data processing, model integration, retrieval, system architecture, user interfaces, evaluation, automation and the engineering required to connect AI with real world systems.

## Evidence
### SudoChat: Generative AI, RAG and AI agents

SudoChat is my current proof of concept for exploring how a controlled AI assistant can operate using a defined knowledge base rather than relying entirely on the general knowledge of an underlying language model.

The project explores:

* Copilot Studio
* large language models
* retrieval augmented generation
* knowledge grounding
* AI agents
* tool calling
* evidence based responses
* source attribution
* fallback behaviour
* hallucination reduction
* prompt and instruction design
* responsible AI boundaries
* conversational user experience

A central design principle is that the assistant should not simply generate an answer because a user asks a question.

It should determine whether sufficient evidence exists in its available knowledge before making claims.

This makes SudoChat particularly useful as an experiment in the types of controls that would be necessary for AI assistants operating in higher trust environments.

### OCR and document intelligence at Xaana.AI

At Xaana.AI, I worked professionally on AI based document processing.

I developed and improved pipelines using technologies including PaddleOCR, OpenCV and other document processing components to extract structured information from documents such as invoices.

The system involved more than performing OCR.

The wider pipeline had to:

* accept document inputs
* convert and preprocess documents
* prepare images for OCR
* recognise textual information
* extract useful data
* handle batches of documents
* pass outputs into downstream processes
* operate reliably enough to support a real workflow

My work reportedly reduced invoice processing time by approximately 40 percent while replacing expensive proprietary components with suitable open source technologies that substantially reduced licensing costs.

This gave me professional experience with the practical difference between an AI demonstration and an AI system that needs to process real documents repeatedly.

### SudoSpeed: Computer vision

SudoSpeed is an open source computer vision project designed to recognise Australian speed signs.

The project explores how visual information captured from a camera can be processed and converted into useful information for another system.

SudoSpeed was not designed only as a standalone demonstration.

I made the project available so that other developers and engineers could potentially learn from it, improve it or incorporate speed sign recognition into their own projects.

The technology also became part of MotorHUD.

### MotorHUD: AI integrated into a physical system

MotorHUD combines software, computer vision, embedded hardware and human centred interface design in a motorcycle safety project.

Its objective is to provide navigation information and overspeed warnings through a transparent display while reducing the need for a rider to repeatedly look away from the road.

SudoSpeed provides the computer vision component for recognising Australian speed signs.

The wider architecture also considers:

* camera input
* computer vision processing
* rider speed information
* navigation information
* communication between devices
* embedded hardware
* display systems
* latency
* user attention
* safety

MotorHUD is important evidence because it shows me thinking beyond the AI model itself.

The computer vision algorithm is only one part of a larger engineered system.

### Machine learning and AI algorithms

My formal engineering education included substantial work in artificial intelligence and machine learning.

I have implemented and studied approaches including:

* artificial neural networks
* convolutional neural networks
* recurrent and LSTM networks
* reinforcement learning
* decision trees
* random forests
* support vector machines
* linear and logistic regression
* clustering
* genetic algorithms
* fuzzy systems
* probabilistic pattern recognition
* computer vision algorithms

This included both mathematical study and practical implementation using Python, PyTorch and related tools.

### AI traffic optimisation

As part of my engineering research, I explored intelligent traffic light control using reinforcement learning.

The project used SUMO traffic simulation and investigated a DQN based multi agent approach where traffic signals could respond dynamically to traffic conditions rather than operating entirely through predefined timing schedules.

The research considered areas including:

* reinforcement learning
* multiple intelligent agents
* traffic simulation
* state and action design
* reward function design
* comparison with conventional traffic light control
* how intelligent control affects system level traffic behaviour

This project gave me experience thinking about AI as part of a dynamic environment where actions taken by one intelligent component can affect the wider system.

### Responsible government chatbot architecture

I have also worked with a behavioural scientist while exploring privacy preserving architecture for Australian government chatbot scenarios.

The work considered situations where citizens could unintentionally provide information such as:

* Medicare numbers
* Centrelink CRNs
* addresses
* banking information
* personally identifiable information
* other sensitive government related information

Rather than considering only what an AI chatbot could answer, I explored what the system architecture should do when sensitive information enters the conversation.

This included considering how information could be identified, restricted, protected or prevented from unnecessarily reaching components of the AI architecture.

The work strengthened my view that responsible AI controls need to exist at an architectural level rather than depending entirely on users behaving correctly.

### AI concepts for illegal dumping detection

While participating in DXC's work investigating illegal dumping in the Northern Territory, I also considered how AI could potentially support detection and response.

Importantly, the AI concept came after field investigation.

I travelled to Alice Springs and participated in environmental surveys, data gathering and community engagement before considering where technology could potentially contribute.

Potential AI applications included analysing available information to identify areas of concern and assisting with more targeted investigation or intervention.

The concept should be represented as exploration rather than a deployed AI system.

Its relevance is in demonstrating my approach to AI:

**understand the problem first, then determine whether AI adds value.**

### Chatbots and conversational systems

I have also developed and experimented with chatbot systems outside SudoChat.

My experience includes understanding how conversational interfaces can connect users with information, automation and backend systems.

My interest in chatbots has increasingly developed toward more sophisticated AI agent architectures, particularly where the assistant can retrieve information or interact with controlled tools rather than simply generate conversational text.

### Local models, agents and tool calling

I have independently explored modern generative AI engineering concepts including:

* local language models
* Ollama
* LangChain style orchestration
* retrieval augmented generation
* tool calling
* AI agents
* model context protocols and MCP concepts
* structured outputs
* document retrieval
* controlled system access
* sandboxed experimentation

My focus has been on understanding how these components fit together into useful systems rather than treating the language model as the complete application.

## Breadth of my AI experience
My AI work can broadly be grouped into several areas.

**Perception**

Computer vision systems such as SudoSpeed and OCR systems at Xaana.AI interpret images or documents.

**Prediction and optimisation**

Machine learning and reinforcement learning projects explore how models can classify, predict or optimise behaviour.

**Language and knowledge**

SudoChat and my chatbot experiments explore LLMs, RAG, retrieval, grounding and conversational AI.

**Agents and automation**

My newer work explores AI systems that can use tools, retrieve information and participate in controlled workflows.

**Physical systems**

MotorHUD demonstrates how AI can become one component inside a larger cyber physical engineering system.

**Responsible AI**

My government chatbot research and SudoChat experiments examine privacy, hallucination, evidence, data boundaries and human oversight.

That breadth is important because I do not see AI as one particular technology.

An AI solution may involve a language model in one project, computer vision in another, reinforcement learning in another, or no AI at all if conventional engineering provides a safer and more reliable solution.

## Relevance to the Federal Courts
The Court AI Technologist role is particularly suited to this breadth.

Many potential Court AI applications are unlikely to consist of a single model operating independently.

For example, a Court knowledge assistant could involve:

**User interface → identity and permissions → retrieval → authoritative Court knowledge → language model → grounding checks → response → citations → logging and evaluation**

A document processing application might instead involve:

**Document → preprocessing → OCR → classification → metadata extraction → validation → workflow → human review**

An AI agent could introduce another layer:

**User request → agent reasoning → permission check → approved tool → enterprise system → result → validation → user**

I have worked with technologies or concepts across each of these kinds of architectures.

My systems engineering background also means I naturally ask what surrounds the AI model.

For a Federal Court system, I would consider:

* Where does the information originate?
* Is that source authoritative?
* Who is permitted to access it?
* What information reaches the model?
* Where is the information processed?
* What happens when retrieval fails?
* What happens when the model is uncertain?
* Can the system take actions?
* Which actions require approval?
* How are outputs evaluated?
* How are interactions audited?
* What happens when the system is wrong?
* Where must a human remain responsible?

That wider understanding is one of the main strengths I would bring to the role.

## Limitations or gaps
Not every AI project I have worked on is a production system.

Some are research projects, prototypes, university projects or personal experiments.

SudoChat is currently a proof of concept and should not be represented as a large scale production AI deployment.

The Northern Territory illegal dumping AI concept was exploratory and should not be represented as a deployed detection system.

I have implemented and studied many machine learning algorithms academically, but academic implementation should not be presented as equivalent to operating those models at enterprise production scale.

My deepest professional production oriented AI experience comes from my document intelligence and OCR work, while my generative AI and agent experience is more recent and has been developed through research, experimentation and proof of concept development.

This distinction should remain clear.

## Useful links
My portfolio:
https://mustafa-siddiqui.com/

GitHub:
https://github.com/sudoqui

LinkedIn:
https://www.linkedin.com/in/mustafa-siddiqui-32ab73161/

SudoChat project and repository

SudoSpeed repository

MotorHUD project

Xaana.AI AI engineering experience

AI traffic optimisation research

Responsible AI and government chatbot research

Northern Territory illegal dumping project evidence

University AI and machine learning coursework

## Do not claim
Do not claim every AI system listed was deployed into production.

Do not claim SudoChat is a Federal Court system.

Do not claim the illegal dumping AI concept was deployed.

Do not describe academic machine learning implementations as enterprise production deployments.

Do not claim I trained a foundation language model from scratch.

Do not claim I created PaddleOCR, OpenCV, Copilot Studio, Ollama, LangChain or other underlying technologies.

Do not claim I have several years of production generative AI experience.

Do not claim I have deployed autonomous AI agents with unrestricted access to government systems.

Do not claim every project used AI simply because I worked on it.

The accurate representation is that I have worked across a broad spectrum of AI, from classical machine learning and computer vision through professional document intelligence to modern LLMs, RAG and AI agents, with increasing focus on how those technologies can be deployed responsibly as components of wider engineered systems.