A creative technologist

Hasan Alavi  • Peerlist

This study presents a potentially groundbreaking training strategy for neural networks, with the objective of achieving authentic multimodal functionality and improved generalizability. Inspired by the remarkable versatility of the human brain and the multifunctionality exhibited by human organs, such as the tongue’s capacity to perceive temperature, taste, texture, and humidity, we propose a novel training approach and architecture. Diverging from traditional task-specific models and existing multimodal models—wherein multimodality is achieved by connecting multiple models or utilizing deep autoencoders that primarily consider cross-modality learning during feature extraction while utilizing a single modality for supervised training and testing.

Novel Training Algorithm to Multimodal Functi

Leveraging the power of YoLo-v7 algorithm, this project focuses on aircraft localization and classification using satellite imagery. By harnessing advanced object detection techniques, it enables accurate identification and categorization of aircraft. This project has future potential for applications such as aerial surveillance and monitoring, contributing to  improved airspace security

Aircraft class localization and classificatio

In medical AI, even a highly accurate model (e.g., 95%) carries cumulative error, meaning a small fraction of predictions can be critically incorrect. When deploying such models in real-world healthcare applications, trust in predictions is paramount. How do we ensure reliability when a model might perform unacceptably on certain cases? To address this, we developed a Confidence of Predictions (CoP) framework, inspired by how humans assess confidence based on experience. Our approach doesn’t just provide predictions, it flags cases requiring human intervention, ensuring that potential misclassifications don’t compromise patient outcomes. This proprietary innovation strengthens AI-driven healthcare workflows, making deep learning models more accurate but also reliable and accountable, a crucial step toward AI-assisted diagnostics with human-in-the-loop validation.

Confidence-Aware Deep Learning Segmentation: 

A simple FM transmitter. Purpose was to understand how any audio signal can be translated on a certain frequency band. It consists of a coil (for setting frequency) and some basic circuit components

FM transmitter

In this project, I developed a PID controller using Arduino and programmed it to minimize different factors like overshoot, settling time etcetera by setting appropriate gains (K’s)

PID-Controller Using Arduino-UNO

Mentored an intern on an exciting project. We used deep learning to help robots navigate unfamiliar environments, even if they're suddenly relocated. Our approach involves a dual-encoder, single-decoder model that combines global maps with real-time lidar data. The goal is to teach robots to recognize patterns from real-time lidar data to map their position on the global map.

Simultaneous Localization and Mapping in Kidn

As a sub-project within the domain of weather forecasting, this research focuses on cloud classification. By leveraging Convolutional Neural Networks (CNN), it advances the accuracy of cloud classification, an essential component of weather prediction and meteorological analysis. The project aims to enhance our understanding of cloud formations and their impact on weather patterns, ultimately improving forecast precision.

Cloud classification using CNN (sub-project o

Designed and implemented an innovative organizational structure called the POD (Product Oriented Delivery) System at OIRRC (Ocular Imaging Research and Reading Center) to address critical operational bottlenecks and accelerate the company's transformation into an AI-first SaaS organization. Key Components: •	Multidisciplinary Teams: Created autonomous units comprising specialists from DevOps, Clinical Operations, Engineering, AI/ML, Finance, and Customer Success •	Problem-Focused Approach: Each POD operates independently, concentrating on specific, clearly defined operational challenges rather than traditional departmental silos •	Agile Structure: Established a flat hierarchy with POD System Leader managing multiple pods, while Department Heads serve as strategic advisors and subject matter experts Strategic Objectives: •	Accelerate MVP development and deployment cycles •	Enable autonomous decision-making to minimize centralized bottlenecks •	Reduce operational inefficiencies in manual processes (site onboarding, certification tracking, image submissions) •	Foster enhanced cross-functional collaboration and innovation Implementation Framework: •	90-Day Pilot Program: Structured rollout with clear phases (team formation, training, experimentation, evaluation) •	Performance Metrics: Tracked time-to-MVP, decision autonomy frequency, error rate reduction, and collaboration effectiveness •	Problem Portfolio: Addressed critical issues including automated quality control, certification processes, database optimization, and AI-driven image analysis Outcomes: •	Significant reduction in turnaround times for certifications and onboarding •	Enhanced data accuracy with fewer compliance-related CAPAs •	Improved scalability without proportional increases in human resources •	Foundation for advanced AI integration in clinical trial workflows

Implementation of Cross-Functional POD (Produ

Combining the knowledge of embedded systems, programming and IoT, this project was a complete IoT-system based solution. It includes different serial and wireless communication and embedded system approach

IOT System and Architecture Using ESP-32

The project’s primary focus was automating Optical Coherence Tomography (OCT) layer segmentation, specifically targeting ILM (Internal Limiting Membrane) and BM (Bruch's Membrane) layers. These advancements enabled precise extraction of retinal thickness (fovea, para-fovea, and peri-fovea) metrics and the estimation of C-scan volumetric data, both critical for comprehensive eye disease diagnostics and treatment planning. Designed a custom U-Net architecture, tailored for OCT imaging, which incorporated an adaptive learning rate strategy to effectively avoid local minima during training. The network design included an attention mechanism, enabling the model to focus on relevant retinal regions. Advanced deep learning techniques were employed to address challenges such as imbalance in training data and variability in image quality. The introduction of novel masking techniques further enhanced the pipeline, ensuring consistent preprocessing and improving segmentation outputs for detecting diseases like Dry AMD, Wet AMD, Edema, and Geographic Atrophy. Key contributions include: Disease Detection Models: Developed and fine-tuned deep learning pipelines for detecting Dry AMD, Wet AMD, Edema, and Geographic Atrophy, achieving a 3x improvement in model performance. Tech Stack Evolution: Transitioned codebases from TensorFlow to PyTorch in collaboration with peers, enhancing flexibility and compatibility. Pipeline Optimization: Streamlined the entire AI workflow from data preprocessing to post-processing, introducing novel masking techniques and architectures for a 2x performance boost on the same dataset. Compliance Training: Completed GDPR and HIPAA compliance training to align AI development with regulatory standards.

Ophthalmic Diagnostics Using Deep Learning: I

This project consists of a lock system which ensures safety. It was implemented using basic logical gates and IC’s.

Digital Combinational Lock

This project was based on Arduino UNO, it calculated reactive, real and apparent power of the connected/applied load.

Reactive, Real and Apparent Power Meter

Built a full-stack strategic management platform that solves a problem most growing organisations quietly struggle with: leadership losing visibility as work scales across teams, projects, and layers of delegation. BEV gives decision-makers a single command centre where nothing falls through the cracks. By structuring work across a 6-level hierarchy — Strategy → Sub-Strategy → Goal → Sub-Goal → Task → Sub-Task — it creates a direct, traceable line between high-level organisational intent and the ground-level work being executed daily. Every task exists in context. Every bottleneck surfaces up. The impact is clarity at scale. Managers no longer need to chase updates across tools, threads, or meetings — progress is aggregated automatically up the hierarchy, workloads are visible per member, and role-based access ensures every stakeholder sees exactly what's relevant to them without noise. Threaded comments with @mentions keep conversations anchored to the work itself, eliminating the disconnect between communication and execution. Five purpose-built views — flowchart, list, Kanban, Gantt, and dashboard — mean leadership, project managers, and individual contributors all get the lens that fits how they think, without switching tools. The multi-database architecture supports isolated environments per team or client, each with its own credentials, making it production-ready for organisations managing parallel workstreams. The result is a tool that turns strategic planning from a document into a living system — one that holds teams accountable, keeps leadership informed, and makes the gap between strategy and execution visible enough to actually close it.

BEV: Bird's Eye View: Strategic Management Co

This project compares three gait analysis methods: Qualisys QTM (considered the gold standard), OpenCap software that analyzes smartphone videos, and AI models that identify body points from videos. The aim is to see if OpenCap and AI models can be as accurate as QTM, and how can we make something more reliable than Qualisys QTM. Potentially making gait analysis more accessible and affordable.

A Comparison of Gait Analysis Using Qualisys 

Developed a biomechanical analysis system using Physics-Informed Neural Networks to predict Ground Reaction Forces and moments from raw body-marker data captured via Qualisys motion capture system. The project addressed a significant constraint in biomechanics research at BERG-NUST by enabling accurate force estimation without force plates for inverse dynamics calculations. The model combines LSTM networks with physics-constraint layers enforcing momentum conservation and gravitational principles, achieving a test RMSE of 22.49 and MAE of 8.78. This innovation enhances accessibility to biomechanical analysis for applications in sports science, rehabilitation, and clinical diagnostics by eliminating the need for expensive force measurement hardware while maintaining analysis accuracy. In collaboration with a research mate at BERG-NUST (https://sites.google.com/view/biomedicalresearchgroup/team?authuser=0).

BioForce PINN: Physics Constraint GRF Estimat

This project dives into a comprehensive examination of the various factors influencing post-implant surgery outcomes. By conducting analysis using system dynamics, it aims to shed light on the contributors to patient recovery and surgical success. Link to Thesis: http://10.250.8.41:8080/xmlui/handle/123456789/42880

Identification & Analysis of Factors Impactin

Variations among Optical Coherence Tomography (OCT) devices, such as Topcon, Spectralis, and Cirrus, often lead to inconsistencies in retinal B-scan segmentation outcomes. This project aims to create a novel model that maintains accuracy across different imaging platforms. Drawing inspiration from natural patterns and mathematical principles, and employing techniques like adversarial feature blending, data augmentation, and generative adversarial network-based domain adaptation, the model seeks to overcome device-specific limitations. The goal is to ensure reliable and precise segmentation results, enhancing the integration of such models into diverse clinical settings.

Developing a Device-Agnostic Model for Consis

In this project, we are leveraging Optical Coherence Tomography (OCT) scans to develop a deep learning-based segmentation framework for Intra-Retinal and Sub-Retinal Fluid (IRF & SRF) analysis. This segmentation is crucial for assessing disease severity, aiding clinicians in making informed treatment decisions. Beyond achieving high accuracy, we are actively working on proprietary advancements that enhance our model’s ability to generalize across different imaging modalities and devices. Medical imaging varies significantly based on the equipment used, and segmentation models often struggle when applied to scans from devices they weren’t trained on. Our novel techniques aim to bridge this gap, ensuring robust performance even on unseen data sources, a key step toward real-world clinical adoption.

Deep Learning-Based Intra-Retinal and Sub-Ret

Home automation demo using google firebase as dashboard and NodeMCU as a controller through API to enable working with personal voice command.

Home automation using Firebase+NodeMCU with G

Using a micro controller Atmel 8051(assembly language), this device protects any electrical appliance from over and under voltage. This task was achieved by using mapping and interpolation techniques for voltage divider, which concludes the changing/maintaining of current state of the circuit.