ai/ml engineer · multan, pk

Moiz
Baloch

A Computer Science undergraduate building machine learning, deep learning, and LLM systems, from experimentation to production. Fine tuning models, serving them, and engineering the infrastructure around them.

$fine tuning llms with qlora
01About

From Flutter apps to intelligent systems.

I started building production mobile and full-stack applications with Flutter, Node.js, and Firebase. That engineering foundation taught me how to ship real products. It became the base for the next step of the stack.

My work now centers on machine learning, deep learning, and LLM engineering: fine tuning models with QLoRA, building computer vision systems, serving them through FastAPI, and engineering the deployment layer around them.

The strongest thread across my work is systems thinking: moving from an idea to a client, a backend, a model, an API, and a deployment strategy. I build complete AI systems, not isolated notebooks.

66+
public repositories
2
production roles
1
published paper
3
LLMs fine tuned
Stack evolution
01

Application engineering

Flutter · Dart · Firebase

02

Full stack systems

Node.js · MongoDB · Socket.IO

03

ML engineering

XGBoost · LightGBM · SHAP · Optuna

04

Deep learning & vision

PyTorch · YOLO · OpenCV · ByteTrack

05

LLM engineering

Transformers · LoRA · QLoRA · TRL

06

AI systems & serving

FastAPI · Docker · MLflow

07

Model internals

AtlasLLM · AtlasMoE · inference optimization

02Mobile era

Before the models, there were apps.I built a lot of them.

I didn't start in machine learning. I started by shipping apps. My first serious engineering era was mobile development with Flutter, and I treated it the way I treat models now: build it, ship it, make it real.

Full applications, not screens. Clients in Flutter, backends in Node.js, MongoDB databases, real time chat over Socket.IO, Firebase for auth and sync, and admin panels to run it all. The same product shaped for phones, tablets, and the people using them.

That era is why AI never felt like a jump. When I moved into machine learning, I already understood the whole product around the model: how a user touches it, how data flows, how a backend serves it. The intelligence layer just became the next thing to build.

FlutterDartFirebaseSupabaseMongoDBNode.jsSocket.IOREST APIsFVMAndroid
01

AL-Safeena

Service marketplace and FYP. Flutter client, Node.js backend, real time chat, MongoDB, React admin panel, ML assisted provider ratings.

02

ShopEase

Ecommerce application with a complete shopping flow.

03

Synk

Chat application built on real time messaging.

04

iWENT

Event oriented application.

05

QuickPDF

PDF utility application.

06

BLYND

AI oriented social application concept.

07

UXelerate

UI and UX focused project work.

08

Khushi Kids

Children's story and educational app, designed in Figma.

09

Trade With Shaw

Built during the software development phase.

10

IdeaCatalyst

Web project built for Air University.

ten builds across the era, each with its own stack and its own problems

03Skills

The stack I build with, end to end.

From model training and fine tuning to serving, deployment, and the infrastructure that keeps AI systems running.

ML frameworks

//
PyTorchTensorFlowHugging Facescikit-learnXGBoostLightGBMCatBoost

Deep learning & vision

>>
CNNsEfficientNetYOLO11OpenCVByteTrackMediaPipe

NLP / LLMs

~
TransformersPEFTLoRAQLoRATRLBitsAndBytes

MLOps & serving

$
FastAPIDockerMLflowDVCGitPostgreSQL

Languages & tools

>
PythonSQLDartJavaScriptC++RustFlutterNode.js

Currently exploring

?
KV cachingquantizationMoE routinginference enginesFlashAttentionRAG
04Selected Work

Projects I've built & shipped.

Model fine tuning, computer vision, and AI systems, each with the engineering detail behind it. No screenshots, just the work.

~/projects/clinical-digital-twinCompleted

Clinical Digital Twin

2024/26

AI-powered personalized patient simulator trained on 534k+ MIMIC-IV hospital admissions. Five calibrated risk models predict mortality, readmission, ICU admission, length of stay, and clinical deterioration from the first 24 hours, with a grounded RAG assistant that refuses rather than fabricates.

LightGBMXGBoostFastAPIReactSHAPRAG
GitHubview code
~/projects/vision-guardCompleted

VisionGuard

2026

Commercial-grade computer vision platform analyzing live webcam, USB, RTSP, and IP streams in real time, with YOLO11 detection, ByteTrack tracking, and an event engine with nine rules served through a FastAPI + WebSocket API.

YOLO11ByteTrackOpenCVFastAPIWebSockets
GitHubview code
~/projects/researchmateCompleted

ResearchMate

2026

A 3B instruction model fine tuned with LoRA on PubMedQA/arXiv so it can answer research questions from scientific literature. Evaluated on latency, quality, and hallucination behavior.

Qwen2.5-3BLoRATRL4-bitPubMedQA
GitHubview code
~/projects/math-tutorCompleted

Math Tutor

2026

Qwen2.5-1.5B-Instruct fine tuned on GSM8K with QLoRA to solve math problems step by step. Trained on a Kaggle T4 with 4-bit quantization and published to Hugging Face.

QLoRAGSM8KPEFTTransformers
GitHubview code
~/projects/rbc-anomalyCompleted

RBC Anomaly Detection

2024/25

End to end medical diagnostic pipeline detecting red blood cell anomalies with XGBoost (~92%+ validated accuracy) and SHAP explainability, plus LLM generated clinical reports with a rule based fallback for zero downtime.

XGBoostSHAPOllamaQwen2.5-7B
GitHubview code
~/projects/visionforgeCompleted

VisionForge

2026

Real time, gesture controlled augmented reality effects. Hand and pose tracking drive visual overlays on live camera input.

OpenCVGesture ControlAR
GitHubview code
~/projects/jarvisIn Progress

J.A.R.V.I.S

2026

Local first, privacy focused voice assistant: Whisper speech-to-text, Qwen2.5 reasoning, 40+ tool actions, and persistent memory for hands-free automation of repetitive tasks.

WhisperOllamaQwen2.5Python
GitHubview code
~/projects/echo-projectxIn Progress

Echo-ProjectX

2026

Offline capable, file system based automation assistant. The engineering foundation for a J.A.R.V.I.S style agent that runs without the cloud.

PythonFile SystemsAutomation
GitHubview code
~/projects/blackholelabIn Progress

BlackHoleLab

2026

A C++ computational physics simulator exploring black hole dynamics, particle trajectories, photon paths, gravitational lensing, and relativistic effects.

C++PhysicsSimulation
GitHubview code
05Roadmap

What's next, in the lab.

Forward looking builds I'm working toward. These are honest statuses: planned and in progress, not claimed as shipped.

Planned01

AtlasLLM

A dense transformer LLM built from scratch in PyTorch, focused on understanding and implementing transformer internals rather than only fine tuning an existing model.

priority build
Planned02

AtlasMoE

The successor to AtlasLLM: a Mixture of Experts LLM built from scratch, exploring sparse routing, distributed concepts, and efficient optimization.

after atlasllm
In Progress03

F.R.I.D.A.Y

The next generation voice assistant evolving from J.A.R.V.I.S, with deeper tool use, richer memory, and fully local first operation.

successor to jarvis
Planned04

NeuroScope

A neural network and ML reverse engineering visualization platform. A serious desktop project for exploring models from the inside.

desktop tooling
Planned05

PhysicsLab AI

A scientific simulation platform combining physics, equations, and simulations, connecting AI engineering with scientific computing.

science + ai

> The pattern is deliberate. I keep moving down the stack: from apps to APIs, models, and now model internals.

builder mindset, always
06Experience

Where I've shipped.

Brain Hub Technologies (Trade With Shaw)

Jun 2025 to Oct 2025

AI/ML & Mobile Developer Intern

Multan, Pakistan

  • >As the only AI/ML developer on the team, designed and deployed stock price prediction models using Python based ML pipelines, contributing to a ~30% increase in trading revenue across internal operations and international client portfolios.
  • >Built end to end pipelines from feature engineering to model serving, ensuring consistent, measurable performance across multiple client portfolios.
Pythonscikit-learnXGBoostFastAPIML pipelines

Al-Safeena

Oct 2025 to Jul 2026

Senior Mobile Application Developer

Saudi Arabia · Remote

  • >Built a home services freelance marketplace giving Saudi vendors a centralized platform for consistent business, built with Flutter and Node.js backends, contributing to a ~40% increase in company revenue.
  • >Continuing as final year project: managing incremental feature development and deployment of the live platform.
FlutterDartNode.jsMongoDBSocket.IO
07Education

Where I'm learning.

B.S. Computer Science

Air University, Multan Campus

Oct 2023 to Jun 2027

Final year project: Al-Safeena, the live home services marketplace built for a Saudi Arabian client.

Relevant coursework

  • Machine Learning
  • Deep Learning
  • Data Structures & Algorithms
  • Databases
  • Mobile Application Development
08Certifications

Formally trained.

Jul to Aug 2026

GIKI Advance AI Bootcamp

Ghulam Ishaq Khan Institute (GIKI)

Intensive bootcamp covering modern AI and ML engineering practice, from training to deployment.

2026

Anthropic Claude 101 Courses

Anthropic

Hands on LLM engineering foundations, including Claude Code and model best practices.

09Publications

Research, peer reviewed.

Co-author · 2026

Adoption Readiness and Perceived Reliability of Generative AI Tools in Software Development Education

A. Bilal, E. Qazi, A. Noor, A. Moiz, M.A. Lodhi

Qualitative Research Journal for Social Studies (HEC Recognized) · 3(2), pp. 293 to 320

10Contact

Let's build something intelligent.

Open to AI/ML engineering roles, research collaboration, and interesting projects. If you're building something in the ML or LLM space, I'd like to hear about it.

Quick compose

Open in Gmail WhatsApp me

Nothing is sent automatically. This opens a prefilled draft in Gmail or a chat in WhatsApp.