Why Codistan AI Lab?

Codistan believes in computational excellence and we push our limits to produce outstanding software systems. Codistan Artificial Intelligence Laboratory is a manifestation of our belief, where we put together powerful machines, graphical processing units (GPUs) and top-notch scientists to deliver quality intelligence for computing.

1
Motion Planning using Multi Heuristic Search

Year:  2016
Tags: Roboics, Motion Planning

Path planning and search methodologies are readily applicable for lower dimensional data, e.g 3D space and traditional computing data structures. However, available search algorithms become problematic for high dimensional space, e.g robotic joint movements. We introduced multi-heuristic search methodology into the domain of robot motion planning, which was successfully tested on a 16 joint humanoid robot. The aim of this project is to introduce a generic framework for humanoid robotics to do a multitude of tasks. This publication was featured in IEEE proceedings: http://ieeexplore.ieee.org/document/7759694/.
Read more

cataract
Cataract detection using Image Processing

Year: 2016
Tags: Medical, Image Processing

Eye Cataract is a common disease, which starts developing from a young age and often remains undiagnosed because of expensive equipment. Moreover, it is important to track the spread of Cataract over the span ofthe patient’s lifetime. This project circumvents the use of expensive equipment and need for regular checkups for Cataract patients. Smartphone cameras have incrementally provided us access to high resolution images with lower costs. We brought together healthy and cataract suffering images of an eye and trained a deep neural model that could detect and quantify the spread of this disease. This project was featured in ITU Telecom World 2016.
Read more

imu
IMU Based Gesture Detection

Year: 2015
Tags: Signal Processing, Machine Learning

Gesture detection based on hand motion is challenging due to certain physical factors of sensor resolution, frequency and outlying data. Moreover, complex gestures are hard to identify because of indistinguishable similarity among human gestures. We used a dynamic time warping approach to deal with these challenges. The system was mounted on a hand glove and paired with a smartphone over Bluetooth. The produce aims to relievethe user of physically touching the smartphone for interaction. The product is opensource in nature and available at https://github.com/rizasif/gtlib. The product was featured in DICE Islamabad 2015.
Read more

country wide population
Population Distribution using Delaunay Triangulation

Year: 2016
Tags: Geo Information Systems, Machine Learning

Increased use of geo informatics application like Google Maps demands for a more comprehensive yet efficient approach to study population distribution. One of the purpose includes quick geo tagging for point of interests and study the consumer behavior in order for businesses to take advantage. We utilized open datasets of population distribution and generated an approach based on Delaunay triangulation technique to produce heat maps. These generated heat maps are then utilized on small memory units like smartphones. We tested this approach by creating a digital scavenger hunt application that involved the whole city of Islamabad, Pakistan.This distribution system is unique because of its ability to adapt with the population movement trend within certain geographic locations.
Read more

intelligent reward distribution
Intelligent Reward Distribution System

Year: 2016
Tags: Machine Learning

Customer retention is important for any business, most organizations do that by incentivizing their customers by rewards. However, it is more important to incentivize your customers with the right reward, one that they would want. We created a system that learns user’s interest over a period of time and matches them with the right kind of incentive they would like to have. Moreover, the system takes into account the available rewards and then recommends the most deserving users. Using cosine similarity based recommender system, we predict the outcomes of which reward will bring most gratification score for the user. The system was tested on users playing a scavenger hunt application.
Read more

bid base billing
Bid based Automated Billing System

Year: 2017
Tags: Machine Learning

Impression based billing and bid based advertisement is an in demand feature for all monetizing applications. Just like Facebook advertisement model, in which the advertiser has to pay as much times as the ad is being viewed. The challenges of such a system involves intelligent, transparent and fair preference order for advertisements for the right people. The open nature of this systems brings competition into play and fair order in that case is maintained using relative time and preference swapping. The allocation for limited advertisement slots takes into account the user preference and learns over time what kind of ads the user is willing to see. We used clustering methodologies to predict and weight the preference of the user.
Read more

7
Deep Learning based Content Selection System

Year: 2017
Tags: Deep Learning

Automated content selection is a challenging due to the diversity of data and user preferences. The changing demands by user are mostly based on external conditions like weather, temperatureand geographic terrains. These factors directly influence culture and hence preferences of people. The aim of this project is take into account the environmental conditions of the user and predict a food preference they would most likely be interested in. For example, cold beverages on a clear sunny day and a light meal during break hours. We used an open source data collection to train a neural network that meets the demand of the user by studying their environmental conditions. The choice of neural networks translates into the need for a predictor adoptable to changing human nature, using self-training abilities.
Read more

8
Adaptive Story Game

Year: 2016
Tags: Deep Learning

Traditional story based gameplay is built on pre-made scenarios which are exploited by restricting the number of choices a player can make. Therefore, masking the nature of adoptable gameplay. We digitized all the inputs and outputs of a choice based game in order to train a neural network of how the next challenges must be aligned for the player. The aim is to produce an unpredictable gameplay that is more close to the intelligence level and preferences of the player. The neural network built in this case used an evolutionary approach to build scenarios.
Read more

9
Myoelectric Prosthetic Hand using sEMG and flex sensors

Year: 2016
Tags: Signal Processing, Machine Learning

EMG techniques are popular for prediction of amputated body parts. We used a runtime technique called dynamic time warping for detection of 5 classified hand gestures. Miniaturized sEMG (surface electromyography) electrodes were developed and EMG(electromyography) detection was performed over the forearm. The algorithm was implemented on STM32F4o7 while outputs were observed on a 3D printed DC motor operated robotic hand. The gestures include pointing, grasping and stretching of the fingers. This project was intended to help amputated patients by studying data from healthy people.
Read more