AI · Mobile · Full-StackPassau, Germany

Saad Ullah Khan

I buildmulti-agent AI systems

I build AI systems and mobile apps. For three years I have shipped generative AI in production: LLM agents orchestrated with Temporal, RAG assistants and NLP models, plus the web and mobile apps around them. On Android I write Kotlin and GPU code, and my open-source engine Atmo Engine runs machine learning on the device. I work across the whole stack, from the database to Flutter, React Native and Jetpack Compose. Right now I am also doing a Master in AI Engineering at Universität Passau.

Saad Ullah KhanM.Sc. AI Engineering · Uni Passau
Years building real systems
4+
Years building real systems
Faster compliance reporting
40%
Faster compliance reporting
Real-time GPU effects shipped
14
Real-time GPU effects shipped
01About

A short version
of my story.

Three years of work, a Master in progress, and a lot of interest in the parts nobody shows in a demo.

I build the whole thing, not just the model.

Calling a model is the easy part. The hard part is everything around it: the orchestration that keeps a long job alive, the approval step before an agent acts, the retries, and a database that does not break when the same write arrives twice. I build that layer, and then the web and mobile apps that sit on top of it.

At Automateley I was the only engineer. I built a multi-agent platform for accounting and tax work in Australia, from the first idea to an MVP and then a product in production. Temporal runs the agents as durable workflows. The backend is NestJS, the agent layer is Python, the front end is Next.js, with PostgreSQL underneath and CI/CD to ship it. One agent reads ledger data from the Xero API, uses Gemini to find hidden shareholder loans, and made compliance reporting 40% faster.

Before that I was at Aurora Solutions, shipping full products end to end. A restaurant platform with a React Native app, a Next.js web app and a NestJS backend on PostgreSQL, with sign-in through AWS Cognito. An event management app in Flutter with a Node.js and Express backend on MongoDB and Firebase. And the AI work around them: a RAG assistant grounded in company documents, an agent that haggles with shoppers, and fine-tuned open source language models. Across my AI and production work I have used AWS where it fits: Cognito for sign-in, S3 for storage, and Lambda for model training and other jobs.

On my own time I build Atmo Engine, an open-source live wallpaper engine for Android. It renders 14 effects on the GPU with OpenGL ES 3.0 and Vulkan, finds the main subject of a photo with a model that runs on the phone, and never asks for internet access. It is where my AI work and my Android work meet.

Now I am studying for a Master in AI Engineering at Universität Passau and learning German. It gives me the time to learn the theory behind the things I was already building at work.

  • 4+Years building real systems
  • 40%Faster compliance reporting
  • 14Real-time GPU effects shipped

Education

  1. 2026 - 2028 (expected)

    M.Sc. Artificial Intelligence Engineering

    Universität Passau · Passau, DE

    Learning the theory behind the systems I built at work.

  2. 2019 - 2023

    B.Sc. Computer Science

    FAST-NUCES · Islamabad, PK

    Thesis: V-Storm, a scheduler for Apache Storm that uses machine learning.

Languages

  • EnglishProfessional
  • UrduNative
  • HindiConversational
  • GermanA1, working toward B1

Also

  • TUM Hackathon 2026
  • SOFTEC '22 Coding Competition
  • Professional courses in AI, ML and RAG

Toolkit

Tools I have really used at work and can explain.

  • AI & Agents
    Generative AILLMsNLPTransformersMulti-Agent SystemsAgentic AITemporalLangChainRAGEmbeddingsPrompt EngineeringMCPLLM EvaluationOpenAI APIGeminiHugging FacePyTorchTensorFlowLoRA / QLoRAscikit-learnComputer VisionOn-device ML (LiteRT, ML Kit)
  • Mobile
    KotlinAndroidJetpack ComposeMaterial 3Coroutines & FlowMVVMHiltRoomRetrofitOpenGL ESVulkanFlutterReact NativeDartFirebase
  • Backend
    NestJSNode.jsExpressFastAPIFlaskRESTGraphQLMERNMEAN
  • Frontend
    Next.jsReactAngularTailwind CSS
  • Cloud & DevOps
    AWS LambdaAWS CognitoAWS S3CI/CDGitHub ActionsFastlaneDockerKubernetesPrometheusGrafanaElastic Stack (ELK)Azure ML
  • Databases
    PostgreSQLMongoDBMySQLFirebasePineconeQdrant
  • Languages
    TypeScriptPythonDartKotlinJavaScriptJavaSQLC/C++
  • Developer Tools
    GitGitHubGradlePlay ConsoleLinuxPostmanJIRAAgile / Scrum
02Experience

Where I have
worked so far.

Four jobs, from helping in a university lab to building an AI agent platform on my own.

  1. Lead AI & Full-Stack Engineer

    Apr 2025 - Jun 2026

    Automateley · Australia (Remote)

    The only engineer on an accounting and tax platform, from the first idea to an MVP to a product in production.

    • Built the whole system on my own: the NestJS backend, the Python layer that runs the AI agents, and the Next.js front end.
    • Used Temporal to hand out tasks to the AI agents. Long jobs keep their state, try again after an error, and you can follow every step.
    • Built agents that handle tax work, chase client documents, check Division 7A rules and draft legal text. A person checks every result.
    • Built a Division 7A agent that reads ledger data from the Xero API, uses Gemini to flag hidden shareholder loans, works out the minimum yearly repayments against ATO rates and drafts the loan agreement. It made compliance reporting 40% faster than the fully manual process, with accountants still making the final call.
    • Connected the platform to the ASIC Register API to verify entities and to the ATO Standard Business Reporting API to prepare lodgements.
    • Added approval steps so a person signs off each agent action, checks that skip repeated model calls to save money, and fixed IDs so a repeated write never creates a duplicate.
    • Launched the MVP on Hetzner, then moved the platform to AWS for production, with PostgreSQL as the main database for every client and CI/CD that deploys the whole platform.
    • Temporal
    • Python
    • Gemini
    • NestJS
    • Next.js
    • PostgreSQL
    • AWS
  2. Software Developer

    Jul 2023 - Mar 2025

    Aurora Solutions · Stockholm · Islamabad office, hybrid

    Full products end to end: mobile, web, backend, and the AI features inside them.

    • Built the onboarding assistant and the pipeline behind it. It turns hundreds of policy documents into small searchable pieces in a vector database, so every answer comes from a real document. New hires sent 68% fewer questions to HR.
    • Built the receipt OCR tool from start to finish: read the photo, pull out the fields, check them, and write them into the HR sheet. It reads fields with 82% accuracy and halved the time HR spends on each claim.
    • Wrote the haggling logic for Haggle AI with LangChain and the OpenAI API on a NestJS, PostgreSQL and GraphQL backend, and built the full subscription and payment flow.
    • Fine-tuned open source language models with PyTorch, Hugging Face and LoRA, and prepared the training data: JSONL format, cleaning, removing duplicates, and keeping test data out of training.
    • Tested models with task scores and with a model acting as judge, watched the training curves, and shipped AI APIs with FastAPI and Docker.
    • Built a full restaurant platform: a React Native app, a Next.js web dashboard and a NestJS backend on PostgreSQL, with sign-in through AWS Cognito.
    • Built an event management app in Flutter, with a Node.js and Express backend in TypeScript, MongoDB for data and Firebase for auth.
    • Worked in two-week Scrum sprints and built shared front-end components in Next.js, React and Angular that helped the team deliver 20% faster.
    • React Native
    • Flutter
    • Next.js
    • NestJS
    • Node.js
    • LangChain
    • PyTorch
  3. Associate Software Engineer (Intern)

    Feb 2023 - Apr 2023

    Developers Den · Islamabad, PK

    Trained a Dutch language model that did not exist yet.

    • Trained a Dutch GPT-2 model for a client. There was no good Dutch model at the time, so I did the data preparation, the tokenisation, the training and the testing.
    • Used AWS Lambda to train and serve the model.
    • Built B2B chatbots on custom NLP models for the client.
    • Improved the data cleaning and training, which raised intent recognition accuracy by 25% from the first model version to the final one on our internal evaluation set.
    • Python
    • GPT-2
    • NLP
    • AWS Lambda
  4. Lab Demonstrator, Programming Fundamentals

    Aug 2022 - Dec 2022

    FAST-NUCES · Islamabad, PK

    Helped first year students learn to find their own bugs.

    • Helped students with programming logic, debugging and making their code faster.
    • Wrote lab exercises and marked programming assignments.
    • C++
    • Teaching
03Selected work

Things I have built
and shipped.

Products I built at work, and the projects I built on my own because I wanted to.

At work

07

Open a row to read more.

  • An accounting and tax tool for firms in Australia. AI agents do the tax work, chase client documents, check Division 7A rules and draft legal text, and a person checks the result. I built every part of it on my own: the NestJS backend, the Python layer that runs the agents, and the Next.js front end. Temporal runs the agents, so one agent can chase a document for three weeks, keep going after a deploy, and still stop and wait for someone to approve. The Division 7A agent reads ledger data straight from the Xero API.

    • Next.js
    • NestJS
    • Python
    • Temporal
    • PostgreSQL
    • CI/CD
    • Xero API
    • Multi-Agent AI

On my own

04

Open source, research and personal builds.

Open SourceOngoing

Agent Harness

700+ tests in CI

An agent system that runs in the terminal. It uses a ReAct loop with a limit on steps and a budget for tokens, file tools that can only touch their own folder, and skills that load only when they are needed. It finds tools at run time over MCP across three servers (scheduling, calendar and web search), keeps long term memory in Qdrant with semantic search, and runs smaller agents (researcher, writer, critic) with clear rules about what each one is allowed to do. Every run is measured with Prometheus, Grafana, OpenTelemetry traces in Jaeger, and logs you can replay.

  • Python 3.12
  • MCP
  • Qdrant
  • Prometheus
  • Grafana
  • OpenTelemetry
  • Docker
  • pytest
Bachelor's Thesis2022 - 23

V-Storm

About 44% more throughput than the default Storm scheduler

My bachelor's thesis, built with a team of three at FAST-NUCES. The normal Apache Storm scheduler treats every machine in a cluster as the same, but real machines are not the same. V-Storm looks at how strong each machine is and splits the work to match. It also looks at the shape of the job, the free resources, the network traffic and the type of job. To find the type, it opens the submitted JAR file, reads 22 things from the code, and uses a decision tree to say whether the job needs CPU, GPU or memory. That answer then decides how much the CPU, RAM and GPU of each machine count when picking a slot. Tasks that talk to each other a lot are put on the same machine, to keep the traffic off the network.

  • Java
  • Apache Storm
  • Python
  • scikit-learn
  • MERN Stack
  • Bash
  • Distributed Systems
Personal Project2021 - 23

TechShop AI

A shopping app for tech products, built in Flutter. It pulls listings from different platforms, uses AI to sort them into the right categories, and then suggests the best option for what you actually need, so you do not have to compare ten tabs yourself.

  • Flutter
  • Dart
  • AI
  • Recommendations
  • E-commerce
Open SourceOngoing

Atmo Engine

4,000+ Play Store downloads · 300+ GitHub stars

An open source live wallpaper engine for Android that I build on my own. It renders 14 real-time effects (atmosphere, reeded glass, halftone, sketch) with two GPU backends, Vulkan and OpenGL ES 3.0. A segmentation model runs on the phone to keep the main subject sharp: ML Kit in the Play build, U2NetP on LiteRT in the F-Droid build. The app has no internet permission, so photos never leave the device. The UI is Jetpack Compose with Material 3, and one codebase ships five build flavors through GitHub Actions, which checks every file before it is signed.

  • Android
  • Kotlin
  • Jetpack Compose
  • OpenGL ES 3.0
  • Vulkan
  • GLSL
  • LiteRT
  • ML Kit
04Off the clock

Life outside
code.

A few things I enjoy away from work.

I like coffee more than is probably normal. Espresso, pour over, the whole process. It is a craft where you see the result in a few minutes and not in a few sprints.

Most of my walks are an excuse to take photos, usually of old towns and the buildings in them. Everything below was shot on a phone. The limit makes me look more carefully, not less.

I like to take Android apart. I root my phone, try custom ROMs, and add features from other brands with tools like Pixelify.

And when I want to stop thinking, I play a game for a few hours. Good level design and good software design have more in common than people think.

Shot on a phone

Tap to enlarge

05Contact

Let us build
something.

Hiring, looking to work together, or just want to talk about AI agents or Android? The form below goes straight to my inbox.

Currently

Studying AI Engineering in Passau, and open to what comes next.

© 2026 Saad Ullah Khan

Built with Next.js & Tailwind