Working with PhzMe on adhoc basis and accelerating my hands on backend engineer experience.Working on* Resources Management on Azure Portal* Azure AI Studio deployments* GitHub - CICD* Catalogue Development

Visualax is an incredible platform for developing text-to-3D model/mesh generation solutions. At Visualax, under the supervision of 3D experts, I am working with generative AI models to quickly transform text into a 3D mesh. I am exercising NeRF, NeUS, and depth-based 3D meshes. My day to day tasks involve:1. Researching the top-notch bleeding edge generative models and execute 3D generation PoC.2. Manage AWS EC2 G5 instance, quickly deploy the models for inference.3. Develop FastAPI and streamlit based channels for 3D experts to evaluate the practical performance of AI models.4. Tweak the models to improve the results and incorporate the insights and feedback given by 3D experts into the pipelines.Technologies and tools: AWS, HuggingFace, CiVit, Bash, Python, Pytorch, PIL/OpenCV

I have delivered 25+ computer vision projects on Fiverr. Here is a list of the most highlighted builtprojects.- Built multiple image classification models using CNN, BN, Dropout layers, etc.- Trained object detection models using Yolo, FasterRCNN, DETR (Vision transformer)- Trained GANs for image generation - new face, dress, flower, landscape, and new motif-batik-designgeneration - I mostly utilized Stylegan2-ADA, and DCGAN to complete these projects- Trained GANs for image reconstruction - human face to Simpsons face, image segmentation, imagesuper-resolution - I used CycleGAN, Pix2Pix, and EDSR to complete these projects.- Built a CLI application for Pose classification which first train the Pose-Estimation model using theAlphaPose model to learn the pose of a cricket batsman in the image and this pose is then further fed toa single class SVM classifier to classify the short played by the batsman.- Visa Photo - Built a web application using simple JS and Flask (REST API) that let user upload a faceImage, the system will learn the 68 facial landmarks on it, and then rotate the image so that the nose is vertically aligned. Then used Unet model to replace the background of the face with white, and finally made the file downloadable as well.- Skin Tone Swap - Built a CLI application that removes background from videos and changes the skintone of human from white to dark - provided a range of skin tones that user can choose based on hisrequirements.- Built an application that detected road lanes from the reference of car camera and drew a graph toshow the deviation from the central point of lanesEvaluated these models using Accuracy, Precision, Recall, F1, mAP, FID, SSIM, and PSNR scores.Documented these models, wrote go to manuals and README.md files for easier usage.Tools: Github, Python, JS, Flask, Pytorch, HuggingFace and Tensorflow

I have delivered 5+ projects that mainly work with tabular data. Most of them were predictive modeling projects based on EDA (Exploratory data analysis) and data analytics. Here is a list of highlighted ones:- Conceptualized and trained a 1D CycleGAN network with an encoder-decoder generator; a scratchimplementation to experiment with the improvement of bad PPG signals quality into good quality. Iimplemented GAN loss, cycle consistency loss, and identity loss to control the training of 4 networks inCycleGAN.- Participated in the "American Express - Default Prediction" Kaggle competition where a wide rangeof tabular features, the target was to predict either a particular request will default to return the loan. I did an extensive revision of models, data features, and classifier models and got an accuracy of 78%. Iperformed various iterative steps for data reduction which reduced dataset size by more than 50%.- House Price Prediction - Used person coefficient score for dimensionality reduction and built a unique feature encoding technique to convert string values into numbers, and then I trained a 1-d CNN feature learner with linear layers head to perform regression task. MSE loss dropped from 0.12 to 0.01.- Handled a single DBLP data file of size 2.8GB and read it in a memory-efficient way. Performeddimensionality reduction by using domain knowledge and Pearson coefficient then engineered somefeatures. I also exercised problem formulation and created the target variable. I used this data to train3 machine-learning models. After submitting the documented report and the work, I was awarded fullmarks.

I have delivered 10+ NLP projects via Fiverr and LinkedIn. The highlighted projects are listed below:- Delivered an Android application that used the Chaquopy framework to run Python inside JavaAndroid APK. Given the link of a product from Amazon, AliExpress, Daraz, or Ebay by the user, I usedthe Selenium Python module to scrape the text of reviews under the product and get sentiment analysisby loading a pickled SVM model that I trained in Python. Finally show the sentiment analysis to theuser.- Retrained BERT architecture-based models AraBERT, MarBERT, and Qarib for sentiment analysisand false news detection on a dataset of Arabic Language. I used Hugging Face API in Pytorch anddrew ROC curves to compare the performance of each model on different datasets.- Fine Tuned FinBERT (BERT-based model trained on Financial data instead) to classify the quarterlypublished report of S&P 500 companies into negative, positive, and neutral. Collecting the dataset andlabeling is the most highlighted task.- Fine-tuned BioBERT and built a CLI application which based on the diagnosis report of patients willpredict if a patient will have a short stay or long stay at the hospital.- Built a CLI application that used a BERT-based model BERTopic and generated an extractivesummary

At Oxhain, a renowned cryptocurrency exchange, I am serving as a Data Science Engineer.My role involved:- formulating and purposing the problem- designing and implementing AI models to generate up/down trade signals- handling the version control system using Bitbucket and GithubUtilizing advanced techniques like RNN, LSTM, and XGBoost; I successfully trained these models to:- analyze OHLC time series data- took an active role in deploying these models- ensuring seamless integration- real-time signal generation at every 5-minute and 1-hour interval.Besides using AI systems to generate trade signals, I developed CLI programs that used candle stickpatterns e.g. engulfing patterns, to identify potential moments in the market and create trade signalsaccordingly.

One month internship - My internship covered following jobs:- Configured and settled up deep learning libraries (Cuda, Cudnn) for multi-gpu processing and usedpytorch for multi-gpu model training.- I retrained Multilingual BERT model to make it language specific for Urdu only- I wrote complete pipeline to load data and train models, fine-tuning pipelines to use BERT models fortext classifications using HuggingFace module.Research Fellow (MS Data Science Thesis):- With the collaboration of renowned organization CureMD; I with my supervisor was building anExplainable AI for Predicting if an ongoing regimen (treatment plan) will succeed to cure the currentphase of cancer or it will fail. CureMD provided the machine learning models trained on tabular dataand I used XAI LIME to get feature attribute scores. I improved the stability and consistency of LIME inexplanation generations

Final Year Project - Autonomous Weed Plucker:- I built a prototype of a robot where a microcontroller vehicle having a 6-dof robotic arm mounted on it'stop alongside a camera.- I collected the image dataset from cotton plants field where each image has multiple cotton and weedplants in it; I labeled the dataset using LabelImg (an image annotation tool).- I retrained FasterRCNN with Inception backbone on my dataset.- I wrote Python pipeline that would get live feed from camera mounted on vehicle and pass it to theretrained model to get bound box and class label for each plant in the image. Cotton or weed plant.- Then with the help of my other group members, we converted pixel measurements in the image intoreal world measurements; then used inverse kinematics to find the angles on which we will fold servosand finally wrote Arudino pipeline to write angles on robotics arm so that it reaches weed plants onlyand pluck it.

- Learnt and conceptualized the fundamentals of data science and AI.- Learnt the mathematical model of machine learning algorithms.- Used Python and R to trained different models with reviews dataset e.g. comments classificationunder Youtube videos as a spam comment or original one, Sentiment analysis of products on Googleplay store and data visualizations.