Abubakar Aliyu

Type: Master's Thesis

Goal: Eliminate AI hallucinations in biomedical question answering by architecting retrieval-augmented reasoning with self-correction mechanisms

Role & Team: Project Lead (4 members) - orchestrated comprehensive literature review, engineered LLM finetuning pipeline with quantization, architected Redis/FAISS integration, deployed production-ready HIPAA-compliant system

Tools & Methods: Llama-3.1-8B, SciBERT, Flan-T5, FAISS, Redis, Hugging Face Inference Endpoints, PyTorch, synthetic data generation, LLM-as-a-Judge evaluation

Outcome & Impact: Deployed production-ready system achieving 87% error detection accuracy and 0.82 semantic preservation F1-score, demonstrated effective approach to reducing hallucinations in biomedical QA

Type: AI Research Project

Goal: Fine-tune Llama-3.1-8B to generate creative, brandable domain names while maintaining safety through systematic data curation and evaluation

Role & Team: Individual researcher - created hybrid dataset generation pipeline, implemented systematic quality assessment framework, conducted iterative model improvements through controlled experiments

Tools & Methods: Llama-3.1-8B-Instruct, LoRA fine-tuning, PyTorch, Hugging Face Transformers, Weights & Biases, Claude API, custom model versioning with SHA256 hashing

Outcome & Impact: Achieved 99.1% success rate, 100% safety compliance on test set (334 examples). Improved model quality through systematic dataset curation (reduced generic patterns from 38% to 28%). Tracked experiments in W&B comparing baseline vs augmented models. Implemented model versioning with SHA256 dataset hashing for reproducibility. Deployed live demo with creative domain generation capabilities

Type: Fellowship Project (Centre for Journalism Innovation and Development)

Problem & Goal: In West African newsrooms, journalists face a 15-minute window to update breaking news with credible backstories. Under this pressure, research often comes from unreliable sources. The project aimed to empower journalists with an AI-assisted fact-checking system.

Role & Team: AI Fellow – collaborated with journalists and technologists to design and prototype ChatJourno. Built data collection pipelines, engineered the RAG system architecture, and prototyped the chatbot interface.

Tools & Methods: LangChain, BeautifulSoup, Python, Streamlit, RAG pipeline design, dataset curation from trusted sources

Outcome & Impact: Delivered a functional chatbot that reduces newsroom research time and improves reporting accuracy. Presented ChatJourno at the 2024 Media Innovation West Africa Conference, showcasing responsible AI integration in journalism.

Type: Academic Team Project

Goal: Build a production-ready ML system for JB Link telecom to predict customer churn and reduce the 43% quarterly churn rate

Role & Team: Team Lead (5 members) - architected microservices system, implemented FastAPI model serving, designed automated data pipelines with Airflow

Tools & Methods: Docker, FastAPI, Streamlit, PostgreSQL, Apache Airflow, Grafana, scikit-learn, microservices architecture

Outcome & Impact: Deployed enterprise-grade system with batch prediction optimization, automated workflows, real-time monitoring dashboards

Type: Academic/Personal Project

Goal: Build a fully event-driven, scalable pipeline for processing and analyzing large volumes of transportation data

Role & Team: Individual contributor - architected event-driven ETL pipeline, implemented PySpark transformations, configured fine-grained IAM permissions

Tools & Methods: AWS Lambda, AWS EMR, AWS Glue, AWS Athena, AWS CloudFormation, PySpark

Type: Academic/Personal Project

Goal: Build a fault-tolerant, scalable platform for analyzing student performance data with interactive visualizations

Role & Team: Individual contributor - designed AWS architecture with custom VPC, implemented Streamlit dashboard, configured Auto Scaling Groups

Tools & Methods: AWS (VPC, EC2, ALB, Auto Scaling, S3, IAM), Streamlit, Python data science stack

Type: Academic Project

Goal: Conduct systematic analysis of Stable Diffusion parameters to understand trade-offs between generation quality, speed, and prompt adherence

Role & Team: Individual researcher - designed controlled experiments across multiple parameter dimensions

Tools & Methods: Stable Diffusion, Google Colab, multiple schedulers (DPM, EulerA, DDIM, LMS)

Outcome & Impact: Identified optimal parameter combinations (CFG 15.0, 30-40 steps, EulerA scheduler), established best practices for architectural image generation

Type: Academic Project

Goal: Implement a Generative Adversarial Network to generate realistic handwritten digits from random noise vectors

Role & Team: Individual contributor - implemented generator/discriminator architectures, conducted adversarial training

Tools & Methods: TensorFlow/Keras, GANs, Conv2D/Conv2DTranspose layers, BatchNormalization

Outcome & Impact: Successfully generated high-quality MNIST digits, achieved balanced discriminator/generator loss

Type: Academic Project

Goal: Develop object detection and segmentation systems for trash classification and plant disease diagnosis

Role & Team: Individual contributor - built end-to-end pipelines, conducted systematic model tuning

Tools & Methods: Ultralytics YOLOv8, PyTorch, evaluation frameworks