BS Computer Science
Lahore University of Management Sciences
Expected Jul 2026Lahore, Pakistan
- Dean's Honor List (Spring 2025)
- Coursework across deep learning, generative AI, systems, and algorithms
Mission Log / 2026
Shahreyar Ashraf
AI Engineer and Researcher
I design intelligent systems that move from research insight to production impact in security-critical and high-scale environments.
771 AI-generated pull requests analyzed
125 active blockchain repositories studied
15,000+ repositories mined during filtering
Chapter 02 / Origin
Why I Build
I build products where model quality, software rigor, and measurable outcomes all matter at the same time.
I work at the boundary of machine learning research and real-world delivery. My focus is building systems that are statistically sound, operationally reliable, and explainable under pressure.
From graph-based fraud intelligence and production risk scoring to LLM alignment and anomaly detection, I am motivated by one idea: technical depth should create practical leverage for teams and users.
I approach product and engineering like mission design: define the signal, build the system, instrument it deeply, and iterate until the outcomes are measurable.
Chapter 03 / Journey
Timeline of Growth
Education, research, and production roles converging into one engineering trajectory.
education
Research Publication
MSR 2026 - AI coding agents in blockchain repositories
2026LUMS
- Empirical analysis of 771 AI-generated PRs
- Statistical significance testing across tool performance
AI and Process Strategist
RozeRemit - 365 Care Group
Nov 2025 - PresentHybrid
- Production risk-scoring architecture with graph-enhanced fraud detection
- Real-time scoring, explainability, observability, and CI/CD delivery
Project Head and Developer
AIESEC in Pakistan and AIESEC International
Jun 2024 - PresentLahore and Remote
- Built Nexus operations platform for 800+ active members
- Automated global EXPA workflows with measurable latency reduction
AI Growth Intern
Paismo
Jun 2025 - Jul 2025Lahore, Pakistan
- Built sales automation agents for lead intelligence
- Developed multi-agent newsletter generation system
Data Analyst Intern
Junaid Jamshed
Jun 2023Karachi, Pakistan
- Automated reporting and reduced processing time by 30%
- Used data analysis to improve operational KPIs
Chapter 04 / Missions
Case Studies and Systems
Selected projects with technical decisions, stack choices, and measurable impact.
Mission
Real-time Log Anomaly Detection
A hybrid anomaly detection system for high-velocity logs that combines statistical baselines with transformer sequence modeling.
PyTorchDrain3scikit-learnIsolation Forest
- Reduced false positives with adaptive thresholding in streaming pipelines.
- Modeled contextual event dependencies beyond frequency-only methods.
Mission
LLM Alignment and Optimization
From-scratch language model implementation and preference alignment pipeline with low-rank adaptation for efficient fine-tuning.
PyTorchLoRADPOHugging Face
- Cut trainable parameters by 99% with targeted LoRA adaptation.
- Implemented DPO alignment workflow across the final transformer layers.
Mission
Graph Neural Networks (Temporal and Spectral)
A set of graph-learning systems spanning ST-GCN forecasting, spectral regularization, and policy optimization with GraphSAGE.
PyTorch GeometricGraphSAGEGCNST-GCN
- Improved MUTAG generalization by 7.5% via spectral regularization.
- Modeled traffic dynamics on a 100-node grid with spatio-temporal learning.
Mission
Deep Generative Modeling
A practical generative modeling lab covering DDPM diffusion, CLIP-guided sampling, beta-VAE disentanglement, and PixelCNN.
PyTorchDDPMCLIPVAE
- Built a full DDPM training and reverse-sampling loop from scratch.
- Guided generation with text-conditioned CLIP losses.
Mission
Barter Ease
A cross-platform product exchange app with secure identity, mixed SQL/NoSQL persistence, and scalable cloud deployment.
React NativeAWS EC2AWS CognitoSQL
- Delivered secure identity and transaction flows for exchange operations.
- Optimized throughput for high-velocity exchange requests.
Chapter 05 / Systems
Interactive Skills Dashboard
Filter across libraries, concepts, and production tooling that power my work.
ML and DL Libraries
PyTorchPyTorch GeometricHugging Face TransformersDiffusersscikit-learnPandasNumPyOpenCVMatplotlib
ML and DL Concepts
Transformers (GQA, RoPE)Graph Neural Networks (GCN, GraphSAGE, ST-GCN)Diffusion (DDPM)VAEGANPixelCNNDirect Preference Optimization (DPO)LoRASpectral Graph Theory
Programming and Tooling
PythonC/C++JavaScriptGitDockerGoogle Cloud PlatformAWS EC2AWS Cognito
Web and Application Development
ReactNext.jsReact NativeNode.jsMongoDBSQLGraphQLTailwind CSS
Chapter 06 / Proof
Evidence, Leadership, and Research
Quantitative outcomes and responsibilities that validate the journey.
Achievements
Dean's Honor List
Recognized by LUMS in Spring 2025 for sustained academic excellence.
MSR 2026 Publication
Published research on AI coding agents in blockchain repositories with statistically rigorous analysis.
Nexus Platform Transformation
Led the end-to-end digital transformation for AIESEC operations, supporting 800+ members at national scale.
Leadership
Led nationwide platform modernization initiatives spanning product design, architecture, and delivery.
Coordinated cross-functional workflows across local and international stakeholders.
Scaled systems with observability and governance as first-class engineering requirements.
Impact Metrics
771 AI-generated pull requests analyzed125 active blockchain repositories studied15,000+ repositories mined during filtering1,068 blockchain domain keywords extracted23x faster PR resolution observed800+ members served through Nexus40,000 users across 100 countries supported40% reduction in ticket resolution latency30% reporting-time reduction100k+ graph nodes in fraud simulations
Publication and Research
Studying the Footprints of AI Coding Agents in Blockchain Repositories
MSR 2026 / Published / 2026
Led an empirical study of 771 AI-generated pull requests across 125 active blockchain repositories.
771125Built a multi-source mining pipeline using CoinMarketCap and Alchemy APIs to derive 1,068 domain keywords from 15,000+ repositories.
1,06815,000Applied Kruskal-Wallis and Dunn's tests to show uniform acceptance rates (p = 0.494) but significant velocity differences (p < 0.001).
p = 0.494p < 0.001Found Cursor resolving PRs 23x faster than Claude Code with zero median latency in the sample.
23x
Chapter 07 / Contact
Open a New Mission
If you are building AI products, research systems, or high-trust workflows, I am open to collaboration.