Asad Ayub

AI Researcher & HCI Designer | Building the future of Human-Centered AI

Based in: Lahore, PakistanMost recently: Research Intern, Social Media Disinformation Lab, LUMSExperience: 2 years

Computer Science student at LUMS with a focus on AI, Computer Vision, and Human-Computer Interaction. Proven track record in building 3D reconstruction pipelines, optimizing neural networks for mobile edge devices, and leading user-centered design projects like Project Bol. Passionate about bridging the gap between complex technical research and impactful, user-friendly applications.

Experience

Research Intern

Social Media Disinformation Lab, LUMSLahore, PakistanworkFeb 2026

high-stakes claims verified: 100+high-stakes claims verified

Summary

Independently verified over 100 high-stakes claims regarding the India-Pakistan conflict (2025) during an internship at IDSL, establishing a rigorous primary-source data repository with cross-referenced verification metrics and comprehensive source citations. Analyzed real-time engagement metrics (likes, comments, shares) across 4 platforms — Instagram, Facebook, X, and YouTube — mapping observed spread patterns against a consolidated claims database for cross-platform misinformation analysis.

Results

Independently verified over 100 high-stakes claims regarding the India-Pakistan conflict (2025), establishing a rigorous primary-source data repository.
Analyzed real-time engagement metrics across Instagram, Facebook, X, and YouTube, mapping spread patterns for cross-platform misinformation analysis.
4 platforms analyzed (Instagram, Facebook, X, YouTube)

InstagramFacebookXYouTube

Research Assistant

LUMSLahore, PakistanworkAug 2025

Summary

Benchmarking NAS-optimized neural networks on mobile chips and edge devices.

What I did

Contributed to cross-architecture optimization research under Dr. Murtaza Taj, targeting deployment of NAS-optimized neural networks on mobile chips and edge devices with heterogeneous kernel architectures.
Developed and executed Neural Architecture Search (NAS) scripts across 3+ mobile devices and multiple inference frameworks (ONNX, TensorRT) to benchmark latency, throughput, and memory utilization metrics.

Neural Architecture Search (NAS)ONNXTensorRT

Member & Event Organizer

LAS Society – Adventure Society at LUMSLahore, PakistanworkAug 2024

events executed: 5+events executed
participants managed for major expeditions: 25+participants managed for major expeditions

Summary

Planned and executed 5+ outdoor exploration events, coordinating end-to-end logistics, safety protocols, and participant management for groups of 20+ members. Served as lead organizer for the Andraab Lake expedition, overseeing route planning, camp setup, activity coordination, and safety oversight for all 25+ participants.

Results

Planned and executed 5+ outdoor exploration events for groups of 20+ members.
Lead organizer for the Andraab Lake expedition, managing logistics and safety for 25+ participants.

Skills

Technical

Python

Neural Architecture Search

Computer Vision

Machine Learning

Deep Learning

C++

Hugging Face (PEFT, LoRA)

UI/UX

Figma

Git

GitHub

PyTorch

TensorRT

scikit-learn

OpenCV

SciPy

SQL

ONNX

Data Analysis

PowerPoint

MS Excel

Mern

Word

Typescript

Haskell

Areas of expertise

Research & Factchecking

Team Leadership

Event Management

Projects

TB-ViTAR: Iterative Spatially Grounded Reasoning for TB Diagnosis

Lead Researcher/Designerpersonal

Designed a research framework combining TARA (Think-Act-Rethink-Answer) iterative reasoning with per-step process reward GRPO for tuberculosis diagnosis in chest X-rays using Qwen2-VL-2B fine-tuned PEFT LoRA. Benchmarked 6 configurations — ResNet-50, DenseNet-121, CLIP, SFT, Outcome-GRPO, and Process-GRPO — on the TBX11K dataset (8,400 chest X-rays); top VLM achieved 92% accuracy, 90.8% sensitivity, and mean IoU of 0.273. Demonstrated that process-reward GRPO produced 47% stronger mean training signal and 163% more high-quality training samples versus outcome-only GRPO, with 100% TARA format compliance.

TARAGRPOQwen2-VL-2BPEFTLoRAResNet-50DenseNet-121CLIPSFTTBX11K dataset

View the code

Structure from Motion (SfM) Pipeline

Engineerpersonal

Engineered a complete SfM pipeline in Python featuring SIFT feature detection with Lowe’s ratio test, essential matrix estimation via RANSAC, chirality-checked 3D triangulation, and optimal image-pair selection. Extended the pipeline to incremental multi-view 3D reconstruction using PnP pose estimation, descriptor-based 2D–3D matching, and bundle adjustment (SciPy) with coordinate normalization for improved numerical stability. Exported reconstructed point clouds as RGB-colored PLY files via a modular src package with CLI supporting features, two-view, and incremental reconstruction modes.

PythonSIFTRANSACPnP pose estimationSciPyCLI

View the code

Project Bol

Lead Designer & Researcherpersonal

A voice-first, gamified educational app designed to reduce speaking anxiety and build spoken English confidence for middle schoolers in Pakistan. Developed as a capstone for the CS 466 HCI course at LUMS.

FigmaUI/UX DesignUser ResearchUsability TestingAffinity Mapping

Education

Bachelor of Science

Lahore University of Management Sciences (LUMS)Computer ScienceGPA 3.3 / 4Aug 2023 – May 2027

Coursework

Operating SystemsDatabasesAlgorithmsFoundations of AI/MLNetwork-Centric ComputingComputer VisionDiscrete MathematicsHuman-Computer InteractionFoundations of Generative AIDeep LearningOperations Management

Societies & activities

LAS Society – Adventure Society at LUMS

What drives the work

Objectives: Secure an HCI Research Internship at CHISEL@LUMS to apply HCD principles and AI-assisted development.

Hire AsadReach out about a role, a contract or a conversation.For recruiters

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

More people on Kanvis

Browse more people Create your own Kanvis