Usama Shakeel

GiveBack Portal (AI Automation System)
Led end-to-end AI automation for the Give Back Portal, handling complete document processing and classification, grouping
Purchases, grouping Credit cards workflows.
Implemented secure document categorization using Azure OpenAI, separating ID documents (CNIC, passport, credit cards) from
purchase and general documents.
Fine-tuned Azure custom models for sensitive ID document processing to ensure higher accuracy and data security compliance.
Built a parallel pipeline using Google Cloud Vertex AI (Gemini) for non-sensitive document analysis and structured data extraction.
Developed a secure OCR workflow using Azure models for confidential documents and Gemini for general document parsing.
Integrated Amazon Textract with LLaMA-based processing for redaction handling, and applied additional validation using
Nova/Mistral models for redaction checks.
Engineered and deployed a FastAPI backend on a dedicated server on a custom domain.


Match Report Generator

Developed a match report generation pipeline, processing official match data to produce structured and narrative-rich reports.
Fine-tuned AI models on Google Cloud Platform (GCP) to improve report generation accuracy and maintain storytelling, style, and context.
Performed data cleaning and fact verification, breaking reports into claims and validating each using evidence from match stats, lineups,
and events.
Enhanced contextual understanding by implementing modern embeddings, vector databases (Pinecone/Quadrant), and LangChain
retrievers.
Applied prompt engineering techniques to ensure only inaccurate or uncertain facts were corrected while preserving narrative flow and
descriptive elements.
Resolved challenges like descriptive hallucinations and maintained the richness, emotion, and coherence of rewritten reports.
Tracked match events and report flow carefully, ensuring chronological accuracy and complete narrative coverage

DigitalAalim
Designed and developed a Retrieval-Augmented Generation (RAG) based AI system using OpenAI to answer queries related to the
Quran, Hadith, and Islamic knowledge.
Built a retrieval-driven response pipeline to ensure all answers were grounded in authenticated Islamic sources rather than
generative assumptions.
Implemented a structured validation framework to maintain theological accuracy, consistent referencing, and controlled output
formatting.
Consistent and controlled LLM behavior through Parameters settings, token management, and structured prompt architecture to
achieve stable and reliable responses.
Engineered a scalable FastAPI backend utilizing APIRouters, asynchronous endpoints, dependency injection, middleware layers,
Pydantic schemas, background tasks, and centralized exception handling.
Deployed and maintained the platform for real-world public access, ensuring performance, reliability, and system stability on
(digitalaalim.in)


Medical Chatbot (LLaMA Fine-Tuned)
Fine-tuned LLaMA 2 (13B) on medical conversation datasets to improve domain-specific response accuracy.
Built backend APIs to handle model inference and manage real-time patient-doctor style chat interactions.
Implemented prompt structuring and output filtering to keep responses medically relevant and controlled.
Optimized inference performance by managing token limits and response latency for smoother user experience.
Integrated the model into a FastAPI-based backend with proper routing, validation, and error handling.
Tested and refined responses through iterative evaluation to improve consistency and reduce incorrect outputs.


Automated Arabic Pronunciation Evaluation System
Developed a system to evaluate Arabic letter pronunciation from user voice input.
Fine-tuned Whisper-small for better Arabic speech transcription accuracy.
Processed user audio, converted it into text, and matched it with correct reference pronunciation.
Integrated LLaMA 3 (via Groq API) to analyze differences and generate feedback on pronunciation mistakes.
Implemented scoring logic to measure pronunciation accuracy and highlight incorrect letters.
Enabled real-time feedback so users can instantly see and improve their pronunciation.

deep learning expertise with model design, training, fine-tuning, evaluation, and deployment for vision, NLP, and multimodal tasks.
Built and optimized neural networks and ML models for domain specific tasks.
Extensive experience with pretrained models and transfer learning for custom AI solutions.
Developed end-to-end AI pipelines: data preprocessing, augmentation, model training, and deployment.
Collaborated with MS and PhD researchers to build and optimize advanced AI solutions.
Mentored junior AI developers and contributed to research-level AI projects.