Applied AI
Researcher

MSc candidate at the Federal University of Rio Grande (FURG), Brazil. I investigate the tension between interpretability and accuracy in LLMs and RAG systems, building neuro-symbolic frameworks that ground AI predictions in verifiable institutional evidence.

View Research → Download CV

Femi Samuel Adeola

Applied AI Researcher

FURG, PPGComp — MSc

Rio Grande, RS, Brazil

femi@furg.br

Focus Areas

Research Interests

Investigating how AI systems can be made trustworthy, transparent, and genuinely useful in high-stakes decision environments.

⚖️

Trustworthy & Responsible AI

Algorithmic fairness, bias mitigation, and safety in decision-support systems deployed in institutional contexts.

🔍

Explainable AI (XAI)

Interpretability–accuracy trade-offs, model transparency, and evidence-based explanations for LLM-based systems.

🧠

Retrieval-Augmented Generation

Knowledge-grounded text generation, hallucination reduction, and vector retrieval optimisation at scale.

🔗

Neuro-Symbolic Systems

Combining neural language models with deterministic rule engines to achieve both accuracy and auditability.

Featured Work

PEPG 2.0

A neuro-symbolic RAG system for evaluating all 98 Computing graduate programs in Brazil using CAPES institutional data.

✦ MSc Dissertation — FURG 2024–2026

Grounding LLMs in Institutional Data

PEPG 2.0 combines a deterministic symbolic scoring engine (8 KPIs) with FAISS-backed semantic retrieval and multi-LLM generation (GPT-4o, Gemini 2.5 Flash, DeepSeek). The system transforms black-box LLM predictions into transparent, auditable grade assessments anchored in verifiable CAPES documents — achieving an 87 percentage-point improvement in groundedness over unaugmented baselines.

GitHub Repo

52.6%

Cycle 2 exact accuracy

84.2%

Within ±1 grade (C2)

99.7%

Gemini groundedness

0.800

Recall@10 (FAISS)

100%

GPT-4o / Gemini agreement

Scholarly Work

Publications

Research contributions in trustworthy AI and educational data mining.

IEEE — Submitted

Balancing Accuracy and Interpretability in a Hybrid RAG Framework for Educational Policy — A CAPES Case Study

Adeola, F. S., Borges, E. N., & De Bem, R. (2025)

Phase 1 benchmark (N=150) evaluating GPT-4o, Gemini 2.5 Flash, and DeepSeek across symbolic, prediction, and semantic tasks. Symbolic accuracy 86% across all models; prediction DeepSeek 84%; RAG groundedness 0.6755; Recall@10=0.560. Seven shared symbolic failures traced to data gaps in CAPES records.

Submitted RAG Neuro-Symbolic Educational AI

Computers & Education: AI (Elsevier) — In Prep

Grounding LLMs in Institutional Data: A Neuro-Symbolic RAG Framework for Graduate Program Evaluation and Outcome Prediction

Adeola, F. S., Borges, E. N., & De Bem, R. (2026)

Full population evaluation (N=98 programs, 2 CAPES cycles). Cycle 2: exact 52.6%, ±1 grade 84.2%, MAE 0.697. Grade 7 accuracy 100%. Gemini groundedness 99.7% vs 0% baseline (+99.7pp). Cross-model robustness confirmed — GPT-4o and Gemini identical on 100% of programs.

In Preparation LLMs Trustworthy AI CAPES

Background

Education & Experience

A path from teaching and web development to applied AI research.

Education

Oct 2024 — Present

M.Sc. Computer Engineering

Federal University of Rio Grande (FURG), Brazil

PPGComp — Foundations and Applications of AI

2013 — 2019

B.Sc. Computer Science

Final project: Secondary School MIS design & implementation

2010 - 2013

Diploma in Software Engineering

NIIT

Foundational software development and programming

Experience

Oct 2024 — Present

Graduate Researcher — Applied AI

FURG, GInfo Lab / C3, Brazil

Designed PEPG 2.0 neuro-symbolic RAG system
Built FAISS vector knowledge bases from CAPES data

Mar 2020 — Oct 2024

Web & Data Systems Specialist

Carelifeline Limited, UK (Remote)

Website development and IT infrastructure management
Data-informed insights for operational decision-making

Sep 2010 — Dec 2013

Classroom Teacher

New Discovery High School, Nigeria

Technology-integrated curriculum delivery

Technical Skills

AI & RAG

LangChain LlamaIndex FAISS ChromaDB Embeddings

LLMs

GPT-4o Gemini 2.5 DeepSeek Prompt Eng.

Data Science

Python Pandas NumPy R / ggplot2 Power BI

Tools

Jupyter Git/GitHub Docker GCP Gradio

Get in Touch

Let's connect.

I'm open to research opportunities, research collaborations, and discussions on AI applications in education and institutional policy. Feel free to reach out through any of the channels below.

Download Full CV →