Muhammad Fahad Baig

Developed an end-to-end automated hiring platform, from candidate profiling to onboarding, drastically reducing recruitment lead time.

• The back end was designed using Flask and Docker and deployed on AWS ECR and Fargate. A load balancer was integrated to improve availability, and MongoDB was used as the database back end for NoSQL and unstructured data requirements.
• Built the framework of the system pipeline, from document parsing to information extraction, data normalization, candidate grading, and production deployment.
• Designed a robust PDF parser built on MuPDF that extracts key characteristics from PDF documents, such as text, fonts, indentations, objects, images, and lines.
• Created a heuristics-based prediction and probability model that takes in extracted features and performs heading segmentation and subsequent data population within each extracted heading.
• Trained the spaCy NER model (BERT) on a labeled corpus of education and experience to extract key entities such as organization, dates, roles, accreditations, and grades.
• Designed robust textual libraries using the FlashText algorithm to swiftly extract skills (50,000+ skills and categories), normalize institutions (3000+ institutions), and location data.

Designed and developed an information extraction system to extract personal data from insurance cards, such as name, date of birth, address, insurance plan, codes, copay, and phone numbers, to automate patient registration workflows in hospitals and practices and reduce the time and errors associated to manual data entry.

The entire solution served as an Azure Function deployed to the cloud. An Android application was developed by the Avalon Team that called the back-end function to perform inferences. Solely I orchestrated, devised, and implemented the entire system pipeline, from identifying challenges, finding solutions, and realizing it to a production-grade product. Used OpenCV to identify rectangular card regions from raw images and then applied augmentation and vision routines to correct skew, reduce color variances, and make text prominent. And I also trained the end-to-end optical character recognition (OCR) pipeline using PaddleOCR.

The system achieved a remarkable 93% accuracy on text recognition and detection compared to 55% from Tesseract.

Developed an AI-assisted dialer bot that automatically responds to operator questions, transcribes audio, and identifies hold and pick-up movements. It drastically reduced the time spent initiating a call and waiting to be picked up by a human operator.

I evaluated several open-source automatic speech recognition (ASR) frameworks, such as Vosk, SPINE, OpenVINO, and NeMo, for WER analysis on internal insurance voice calls. Then I designed and implemented a robust messaging-based queuing protocol system on top of RabbitMQ to initiate calls, run transcriptions, detect hold times, cancel calls, and handle interactive voice response (IVR) to process up to 40 calls concurrently.

Developed a comprehensive platform for black-owned businesses to register and verify ownership, enabling them to add their businesses to a growing search engine. This involved building administration, business, and user portals, managing roles and permissions, and creating the entire application stack using React, Python FastAPI, and storage backends like MariaDB, MongoDB, and Elasticsearch. Additionally, an extensive scraping pipeline was implemented to gather data from various sources, including Instagram, while utilizing a wide range of JavaScript and Python libraries for seamless functionality.