Tayyab Nasir

A cloud-based hospital management system (HMS) providing SaaS. The idea was to provide a configurable HMS that resides entirely on the cloud and is accessed by a single-page web application. All the computation is done in the cloud, and only the user interface resides on local machines.

As a back-end developer, I wrote web services to expose data to the application's front-end and researched and developed several proofs of concept to make the application HL7 compliant. I also worked on many end-to-end integrations for effective communication between HealthCloud and other existing systems, such as integrating with Dynamics 365 for syncing patient details between the health management information system and the ERP.

In addition, I optimized the application's code and worked on strategies to improve its performance under heavy loads. I also introduced microservices to separate certain functionality from the monolithic application. I worked on a proof of concept to introduce the use of Cosmos DB to fetch more frequently accessed data faster than the relational database.

An integration-ready software solution for automating the collection of a patient's vital readings, which reduces a nurse's workload, and a web-based portal for managing reading rooms, which communicates with a desktop-based reading location app to capture and process readings from certain devices.

As a full-stack developer, I built the management portal with ASP.NET MVC, Dapper, HTML, CSS, Bootstrap, JavaScript, C#, and jQuery. I also developed the reading room application that fetches vital patient readings from hardware devices and passes them to the server using WPF, SQLite, and Material Design WPF. The application also provides audio instructions to patients utilizing the Windows text-to-speech API.

Additionally, I built a manual nurse auto-ride feature using Opentok WebRTC to provide effective end-to-end audio and video communication between the doctor and the reading room. I also developed an HL7-compliant REST API server to receive patient information from any existing healthcare system to be used with the application.

A cloud-based application for automating commercial and small business administration (SBA) lending processes with a fully integrated set of lending tools. Based in the United States, the application provides functionality to manage small business loan portfolios' origination, processing, closing, and monitoring.

I worked as a full-stack developer on the integration module for the FICO LiquidCredit Small Business Scoring Service, creating individual components for submitting FICO requests and processing the returned business and individual credit scores used for processing loan applications.

In addition, I built credit memo customizations for certain lenders, updating the existing UI controls developed using Kendo jQuery UI components. Using Telerik Reporting, I created the system documents and reports modules for PPP loan forgiveness summaries and PPP credit recommendations. I also worked on the Telerik RAD Controls-based PDF engine for mapping SQL table data to and from the required fields of the PDF forms.

A complete office management portal with features for managing human resources, biometric attendance, employee leaves, workstations, and meeting rooms.

I developed the back-end of the biometric attendance management portal. I built the Windows services to run on the individual attendance terminals handling the biometric attendance devices, to record and sync employees' time in and out, and to break start and end times via thumbprint verification. I wrote REST APIs to receive the attendance records from the attendance terminals and manage them in the MS SQL Server database.

Additionally, I worked as a full-stack developer, building the meeting room and workstation management module using ASP.NET MVC. The module allows viewing, adding, updating, and removing rooms and workstations available on office premises.

I also created the meeting room and workstation assignment and allotment module for use cases involving booking a meeting room, managing attendees, authorizing the start or cancellation of meetings, viewing assigned schedules, notifying start and end times, managing workstations or hot-desking, and notifying workstation assignments.

A system to provide services such as mobile phone top-ups and bill payments using cash or card via an installed terminal, much like vending machines.

I developed the self-service terminal server's back end, writing REST APIs using ASP.NET Web API to provide the required CRUD, authentication, authorization, and terminal sync operations. I also developed the full-stack of the terminal application using WPF, SQLite, Dapper, and gridExtra. I built interfaces for relevant hardware devices such as a payout device, cash validator, card reader, and printer.

Moreover, I conducted the complete research and development of each device involved and worked on the transaction sync manager for a more robust and responsive application, even in network unavailability. I also built the Android application for terminal agents using Xamrin forms for agent verification and terminal maintenance management. Lastly, I created a WPF application for user interactions on the terminal machine and a web portal to manage administrative actions.

A cloud-based application and subsystem of NOVATRAQ that gives borrowers and referral sources a secure environment to conduct all activity required during the lending process for SBA, providing features like customizable SBA loan application wizards, loan application monitoring, and document management.

I developed the loan application interview wizards for SBA, SBA PPP, and SBA PPP forgiveness. I built user controls using Kendo jQuery UI, Backbone.js, and Marionette and created endpoint APIs based on ASP.NET Web API with Entity Framework as the object-relational mapper. I also worked on client-side management for monitoring user activity using SignalR.

Moreover, I built the product match and self-registration modules, which allowed borrowers to initiate SBA PPP and its related forgiveness loan applications. I also created the templates module using Rackspace.net to provide borrowers access to template documents stored at Rackspace.

An application that suggests alternate medication for the provided brand or generic name of the medicine works only on Pakistani products. It enables users to locate nearby pharmacies with the medication in stock. The app also provides pharmaceutical companies and pharmacies with a special interface for managing their products, stocks, and branches.

As a full-stack developer, I built separate ASP.NET MVC and Web API servers and used jQuery and DataTables as primary front-end development tools. I implemented custom, claim-based authorization, and token-based authentication for the Android application and Web API communication. Using the Google Maps API, I developed reverse lookup queries for searching by brand and generic names and the pharmacy locator module.

Finally, I worked on various data scrappers, using Selenium for .NET and various data cleaning and formatting techniques to develop a complete, alternate medicine dataset for Pakistani products. I implemented the manufacturer and pharmacy management modules using ASP.NET MVC.

An open-source project to add server-side filtering, sorting, and aggregation functionalities for Kendo jQuery Grid, supporting .NET, .NET Core, and .NET Standard. I developed the complete package as the project's creator.

A Deep Learning-based end-to-end image classifier to identify different 200 different bird species.

I worked as a machine learning engineer building various models using CNNs and fine-tuning them to achieve the desired results. I worked on developing multiple CNN-based models and tried to fine-tune them using K-fold cross-validation. The best working models involved the Inceptionv3 pertained using the ImageNet dataset. I compiled a greater dataset using multiple existing labeled datasets, including the CUB and NABird datasets.

A complete solution for recognizing printed Urdu text in images, supporting multiple fonts. The research generated a benchmark dataset for printed Urdu using various fonts and Deep Neural network-based models that recognize printed Urdu text lines and words regardless of the font used.

I worked as the researcher, building up a complete set of rationale on what limitations of the current systems the said research overcomes. I scrapped, cleansed, and formatted the data, which was then used to generate the final corpus. I built state-of-the-art CNN and RNN-based DNN models to build a complete OCR system. I worked on the analysis of the dataset and the trained models.

An artificial intelligence (AI)-based resume parser with the ability to provide candidate profiling and suitability score for the provided job description. It provides a complete end-to-end recruitment workflow from job posting to candidate shortlisting.

I led the team in developing the AI-assisted parsing, profiling, and scoring APIs. I designed the code architecture of the Parsing API, implementing the best practices.

The project was developed based on a test-driven development approach using Python's unit test library. I enforced the PEP standards using Flake8. I designed different modules in the application, including the PDF parsing module, resume structure extraction (columns, headings, and sections), resume sort order correction, template-based parsing, scoring, and rating module. I worked on different NLP-based techniques, primarily using a fine-tuned version of Spacy NER to extract specific keywords.
I also led the research and development team in preparing training data for the fine-tuning spacy NER (BERT) model. Also, I led the team in creating an ML model for the classification of resumes using styling and formatting information, using style occurrences as model features.

This system uses deep learning models for face detection and recognition to track patient activities within a clinic, monitoring the patient's entry and exit using video streams. Also, the system is used for timing therapy sessions.

• Worked as the research and development lead, building robust face detection and recognition systems using ArcFace and Facenet512 to create an ensemble model for face recognition and SSD for face detection in live video streams.
• Assisted the development team in making the application architecture for registering patients for face recognition and creating an edge-based detection and recognition system that can communicate more efficiently with the existing monolithic EHR system.
• Overlooked the development of the FastAPI-based monolithic server for dealing with endpoints for recognition.

A platform for building bots that can encapsulate the essence of famous personalities and act on their behalf. The bot can respond to emails and SMS, reply to tweets, or even market or endorse a product on behalf of the personality. Additionally, the bot provided features like voice messages in a tone and voice identical to the personality.

• Worked with a team of machine learning (ML) engineers to explore large language models (LLMs) and how to limit these to the knowledge and style of a particular celebrity.
• Researched different embedding models to extract relevant facts matching the input queries to generate data that can be used for better instruction to tune the LLMs.
• Worked on several optimizations, like using Milvus for faster vector search and exploring open-source LLMs like Llama-V2-7B-chat for lighter and cost-effective LLM solutions.
• Developed an end-to-end FastAPI-based back end to expose the solution and architected multi-tenancy support to the infrastructure for multiple personalities, adding support for configurable workflows for each.
• Worked on integrating text-to-speech and speech-to-text modules to add audio input and output features using Whisper and ElevenLabs.

A Visual Studio code extension that uses our locally hosted LLM for generating code completions, ensuring that the proprietary client code is not exposed to any 3rd-party service.

• Worked as a research consultant with our development team to shape the idea of ensuring client code security within our organization.
• Explored different existing tools and LLMs, building a comparative analysis of the performance and efficacy of each according to our use case.
• Finalized the modification of the existing VS Code extension, tailoring it to our use case and making it configurable to work with our locally hosted LLM.
• Deployed Code Llama-13b on RunPod using the Hugging Face Text Generation Inference (TGI) library and provided API access to the development team.

A system for extracting ingredients used for preparing the items in restaurant menu data. Such data was used to link retailers to restaurants based on the ingredients sold and consumed, identifying target restaurants for retailers.

• Designed a pipeline for applying specific preprocessing techniques for cleansing and normalization of the data.
• Created a 2-step ingredients extractor that, in the 1st pass, tried to extract the ingredients from the provided menu data and, in the 2nd pass, prompted the LLM to fill in the missing ingredients, if any.

The project is currently under development, where an ML-based model is being developed in-house to eliminate LLMs and make a more scalable solution.

A machine learning system for identification of the same product belonging to a particular restaurant existing on different food delivery platforms like UberEATS, Zomato, HungerStation, Mrsool, and Jahez. This duplicate identification was then used to compare prices and find missing products over different platforms.

• Worked as the lead data scientist on identifying and collecting data of interest that can be used to generate helpful features.
• Created a system to extract product attributes from the existing data and used such features to build a semi-supervised tagged dataset of 50,000 examples, which was then trained to train multiple machine learning models to find a similarity score.

This system was used in conjunction with another system first to identify common stores and then compare the products of the store existing on different food delivery platforms. Also, the same system was used to identify and remove duplicate store entries that our crawlers extracted.

The proposed system achieved an accuracy of 98% and was used to process more than 3 million product records.

An add-on for Reki, a platform that provides the easiest way to track recommendations from friends and talk about the shows you watch. The add-on used the power of LLMs to give a way to avoid the cold start problem by recommending movies based on an initial questionnaire.

• Created an end-to-end solution using OpenAI GPT-4 to generate a filter for TMDB to extract movie information using the filter.
• Applied specific NLP techniques to parse the irregular format of the LLM output.
• Experimented with Mistral and Llama models to replace OpenAI GPT-4 usage for reduced scaling costs.
• IWorked on a secure API key-based authentication for communication between the Reki app and our recommender.

A smart machine learning and natural language processing system to extract retail products and their associated attributes for analysis of product sales and competitive market analysis on platforms like UberEATS, Zomato, Hunger-Station, and Jahez.
Worked as the machine learning lead using machine learning and NLP techniques for building an advanced ETL pipeline to identify and extract brands, their associated amount, and other attributes.
Created an optimized pipeline for data cleansing, and normalization by applying NLP techniques like tokenizing, expression evaluation, stemming, and translation to the input unlabeled data.
Worked on optimizing the system by introducing Dask data frames replacing pandas for improved processing time.
Build individual modules for confidence calculation, post-processing, and running a verification cycle based on the associated domain knowledge collected from clients.
The final tagging system was used to tag more than 10 million records extracting the required information.

A machine learning (ML) system for comparing store data from different food delivery platforms to find common stores. This duplicate identification was then used for comparative data analysis and correction of existing tagged data.

• Worked as the lead data scientist on identifying and collecting data of interest that can be used to generate helpful features.
• Created a conventional NLP-based system for tagging around 40,000 examples with humans in the loop.
• Experimented with several ML and DL models for predicting the similarity score using fully connected networks.
• Experimented with Siamese networks to find similarities, but the efficacy of such was not at par.
• Applied geo-fencing to group stores using geo-hashing libraries to reduce the number of matches, improving the overall system efficiency.

The same system was used to identify and remove duplicate store entries that our crawlers extracted. The proposed system achieved an accuracy of 96% and was used to process more than 500,000 records.