Mustafa Abbas

Mustafa Abbas

AI/ML Software Engineer | Research Lead | Full-Stack Developer

Based in: Lahore, PakistanMost recently: Software Engineer - AI/ML & Full-Stack, GoSaaSExperience: 4 years

Hire me

Software Engineer and Research Lead specializing in AI/ML and Full-Stack development. Proven track record in building AI-enabled compliance platforms at GoSaaS and leading novel research in Multi-Agent Path Finding and Urdu ASR benchmarking. Expert in integrating AI agents, microservices architecture, and RAG pipelines to solve complex engineering challenges.

Experience

Software Engineer - AI/ML & Full-Stack

GoSaaSworkJun 2025 – Sep 2026

Summary

Developed and maintained an AI-enabled compliance platform, owning features across the AI, backend, and frontend stack from design through deployment.

What I did

  • Integrated AI agents into the platform to automate data extraction and document interpretation, reducing manual data entry and accelerating compliance workflows.
  • Developed across a microservices architecture spanning Node.js/Express and Java/Spring Boot backend services, with a React-based frontend.
  • Advocated for and drove team-wide adoption of AI-assisted development tools (Cursor, Claude Code, Codex) and the AWS AI-DLC methodology, standardizing workflows that improved development velocity, code quality, and consistency across the team.
  • Collaborated with cross-functional teams to design, implement, and deploy features with a focus on scalability, security, and engineering best practices.

Results

  • Eliminated 100% of manual data-entry effort for suppliers by developing an AI agent that automatically extracts and transforms data from any format.
  • Removed the learning curve for new users by automating the processing of composition data without requiring specific Excel templates.
  • Automated document processing, campaign launches, and identification of non-compliant parts to reduce manual review.
  • Replaced a cache-augmented generation system with a RAG-based chatbot to improve answer quality and add source citations.
  • Advocated for an AI-driven software development lifecycle that increased engineering throughput by roughly 7x.
  • Reduced manual data entry time from 5–20 minutes per part to zero.
  • Reduced AI/context costs by approximately 95% for the user-support workflow.
  • Reduced engineering task delivery time from weeks to days.
AI AgentsNode.jsExpressJavaSpring BootReactMicroservicesCursorClaude CodeCodexAWS AI-DLCRAG

Research Lead - Multi-Agent Path Finding with LLMs

Lahore University of Management SciencesworkSep 2024 – May 2025

Summary

Designed a novel multi-agent pathfinding framework for warehouse environments using a fine-tuned GPT-3.5 controller.

What I did

  • Implemented a feedback loop to parse predicted paths, detect errors, and issue corrective prompts, improving path accuracy and ensuring collision-free navigation.
  • Fine-tuned on datasets derived from BFS (single-agent) and CBS (multi-agent) to adapt to varying warehouse layouts.
LLMsGPT-3.5BFSCBSFine-tuning

Teaching Assistant - CS-370: Operating Systems

Lahore University of Management Sciencespart timeSep 2024 – 2025

Summary

Assisted in a core course with 220 students across two sections, and conducted office hours twice a week to support students with course material and assignments.

What I did

  • Designed, tested, and guided students through complex programming assignments, delivering assignment-based tutorials to reinforce key concepts.
  • Graded exams and assignments, ensuring timely feedback and academic support.
Operating Systems

Skills

Technical

PythonPython
JavaScriptJavaScript
Claude Code
GitGit
Cursor
Hugging Face
NumPy
PyTorchPyTorch
Pandas
Node.jsNode.js
ReactReact
REST APIs
TypeScriptTypeScript
Java
TensorFlowTensorFlow
LangChain
Scikit-learn
Spring BootSpring Boot
ExpressExpress
Codex
Postman
Google Colab
Kaggle
SQLSQL
DockerDocker
HTML/CSSHTML/CSS
C/C++C/C++
Decision Trees
SVMs
GitHubGitHub
Cross validation
LinkedInLinkedIn
RAG
KNNs
Naive Bayes

Languages

Urdu
English

Projects

Landmark 3D Reconstruction

academic
Built an AR Android app to visualize a 3D reconstruction of the Pantheon, using a custom computer vision pipeline with preprocessing, SIFT feature detection, FLANN matching, RANSAC filtering, and Structure-from-Motion to accurately recover camera poses and geometry.
ARAndroidComputer VisionSIFTFLANNRANSACStructure-from-Motion

Model Compression for Deep Learning

academic
Implemented and evaluated pruning (structured/unstructured), quantization (PTQ, QAT across multiple bit-widths), and knowledge distillation (logit, hint-based, CRD) on VGG models with CIFAR-100, achieving significant memory and efficiency gains while retaining accuracy; code and results available in the project repository.
PruningQuantizationPTQQATKnowledge DistillationVGGCIFAR-100

Course Recommendation System

academic
Built a university-wide course recommendation system using a RAG pipeline (LangChain, Mistral-7B, All-MiniLM-L6-v2, ChromaDB), scraping and preprocessing forum reviews, form responses and course outlines to deliver personalized suggestions via a Hugging Face Spaces and Gradio interface.
RAGLangChainMistral-7BAll-MiniLM-L6-v2ChromaDBHugging Face SpacesGradio

Succession Planning Software

academic
Built an ML-driven HR platform using the MERN stack that tracks and analyzes employee performance using KPIs, delivering workforce optimization suggestions, focusing on eliminating biases and enhancing meritocracy and transparency.
MERN stackMLKPIs

Addressing Data Heterogeneity in Federated Learning (FL)

academic
Addressed non-IID data challenges in FL by reducing client drift through latent space alignment between client and global models, using metrics such as KL divergence, Maximum Mean Discrepancy, Wasserstein distance, and L2 norm, achieving state-of-the-art accuracy on CIFAR-10 with label and quantity skews.
Federated LearningKL divergenceMaximum Mean DiscrepancyWasserstein distanceL2 normCIFAR-10FedAvgMMDAdversarial lossGradient harmonizationlatent space alignment

Sentiment and Toxicity Detection

personal
Built a system to detect and classify toxicity into categories like 'toxic', 'severe_toxic', 'obscene', 'threat', 'insult', and 'identity_hate'.
KNNsNaive BayesDecision TreesSVMs

Education

BS in Computer Science

Lahore University of Management SciencesComputer ScienceGPA 3.96Sep 2021 – Jun 2025

Coursework

Advanced Topics in MLGen AIComputer VisionApplied ProbabilityMachine LearningData MiningData ScienceAlgorithmsData StructuresLinear AlgebraCalculus I & II

Recognition

Awards

Graduated with High Distinction

Lahore University of Management Sciences2025

For maintaining a cumulative GPA above 3.80.

Dean's Honor List

Lahore University of Management Sciences2025

For the Academic Years 21-22, 22-23, 23-24, and 24-25.

50% merit scholarship

Lahore University of Management Sciences2025

For being Top 15 in batch ranking for 21-22, 22-23, 23-24, and 24-25.

Patents & publications

WER We Stand: Benchmarking Urdu ASR Models

publication2024

arXiv: 2409.11252 [cs.CL]. Equal contribution by Aamina Jamal Khan and Mustafa Abbas.

With Samee Arif, Aamina Jamal Khan, Agha Ali Raza, Awais Athar

Read it
Hire MustafaReach out about a role, a contract or a conversation.For recruiters

Mustafa's twin is AI, it can make mistakes.

More people on Kanvis

Browse more peopleCreate your own Kanvis