Faruna Godwin Abuh
Applied AI Safety Engineer · Evaluation & Interpretability
Building safety-aware AI systems with focus on model behavior, dataset quality, and real-world deployment
Selected Project Metrics
Exploratory & Demo Deployments · Non-Production Research Projects
Sentences Collected
Deployed
Tested
All tools built for research exploration and learning purposes.
About / Research Philosophy
I focus on applied AI safety — how models behave in real systems, how datasets shape outcomes, and how deployment context introduces risk. My work centers on evaluation, interpretability, and low-resource NLP, where safety issues often appear earlier and more visibly.
Rather than optimizing for scale or performance alone, I prioritize transparency, robustness, and understanding failure modes before deployment. I'm especially interested in how AI safety frameworks apply outside well-resourced settings and how tooling can support safer adoption globally.
Featured Projects
Building AI for Real-World Impact
Red-Teaming LLMs for AI Safety
Automated adversarial testing to identify vulnerabilities in large language models
Systematic red-teaming framework for evaluating LLM robustness against prompt injection, jailbreaks, and unsafe outputs. Implements multiple attack strategies with automated safety scoring.
Igala-English Neural Translation
Production-grade machine translation for low-resource African languages
Fine-tuned mBERT model for Igala-English translation with quality metrics, uncertainty quantification, and interactive demo. Addresses critical language preservation and accessibility needs.
Mechanistic Interpretability Analysis
Understanding transformer attention patterns in low-resource NMT
Deep dive into mBERT's internal mechanisms during Igala-English translation. Visualizes attention heads, analyzes token alignments, and explores how transformers handle morphologically complex languages.
A decoder-only transformer trained on Igala corpus, implementing multi-head self-attention, positional encoding, and custom tokenization. Demonstrates deep understanding of transformer architecture.
An interactive Streamlit application for exploring and auditing low-resource language datasets. The tool provides sentence statistics, token analysis, and linguistic insights to support NLP research and language preservation efforts.
AI Safety & Calibration
Making transformer models transparent through calibration and mechanistic interpretability
Empirical evaluation and mechanistic interpretability analysis of transformer models, with a focus on reliability, calibration, and understanding internal decision mechanisms using GPT-2.
AI Portfolio Assistant
Production-grade conversational AI with memory and intelligent response modes
A full-stack AI chatbot featuring conversation memory, automatic mode detection, real-time logging, and analytics. Built to provide personalized portfolio Q&A experience using Google Gemini API.
Detailed Case Studies
Deep dives into architecture, challenges, and solutions
Safety & Evaluation Writing
Exploring dataset quality, model behavior, and deployment considerations
Project Vision
My long-term goal is to build AI systems that are not only powerful, but also trustworthy and useful in real-world settings. This means designing models that communicate uncertainty, respect local languages, and avoid overconfident or unsafe recommendations.
I'm especially interested in deploying AI where access to expertise is limited, and where responsible design can make a meaningful difference. Whether it's building datasets for underrepresented languages, making models more interpretable, or designing systems that know when to abstain—I believe AI should serve communities, not replace human judgment.
Inclusive NLP
Building for low-resource languages
AI Safety
Transparency and reliability
Real Impact
Serving underserved communities
Let's Connect
Open to applied AI safety, evaluation, and research-engineering collaboration. If you're working on model behavior analysis, low-resource NLP, or responsible AI deployment—let's talk.
💡 Currently exploring: Selective prediction methods, failure mode analysis in low-resource NLP, and production LLM observability