Projects & Experience | Adejare Adelugba

Flagship Research

Final Year Thesis

Flagship · Final Year Thesis

Multimodal Breast Cancer Detection

PyTorch • MedCLIP • TabNet • Early Fusion • Landmark University 2024–2025

Addressed single-modality screening limitations by designing a fusion architecture combining ultrasound imaging with clinical and radiometric data. 87.1% ROC-AUC.

87.1% ROC-AUC 82.7% Accuracy 576-dim Fusion Vector Outperforms Unimodal Baselines

MedCLIP (Vision): 512-dim semantic embeddings from ultrasound images via a vision-language model pretrained on medical data.
TabNet (Tabular): Attention-based processing of clinical metadata — patient age, BI-RADS, tissue density, radiometric features.
Early Fusion: Concatenated 576-dim vector passed through a classification head.
Results: 82.7% Accuracy • 87.1% ROC-AUC — significantly outperforming unimodal baselines.

class CancerFusionModel(nn.Module):
    def __init__(self):
        self.medclip = MedCLIPModel()   # 512-dim
        self.tabnet  = TabNetEncoder()  # 64-dim
        self.head    = nn.Linear(576, 2)
    def forward(self, image, tabular):
        img = self.medclip(image)   # [B, 512]
        tab = self.tabnet(tabular)  # [B, 64]
        return self.head(torch.cat([img, tab], dim=1))

Engineering Decisions & What Didn't Work

Late fusion failed first. Initially trained image and tabular branches independently and combined predictions. Results were worse than either branch alone — the fusion was averaging noise, not learning interactions. Early fusion (concatenating 512+64 embeddings before the classification head) allowed cross-modal feature learning and recovered the gap.
ResNet50 dropped ROC-AUC by ~6 points vs MedCLIP. General vision models don't transfer well to medical imaging semantics. Domain-pretrained encoders are not optional for clinical AI tasks.
Class imbalance caused 94% false-negative rate before correction. Weighted cross-entropy loss fixed this at the cost of ~3% overall accuracy — the correct trade-off in a screening context.

View Repository

Work History

Professional Experience

Internship

Data Engineering Intern

Kontemporary Konsulting Ltd. · Abuja

Aug – Sep 2025

85%

Manual Entry Reduced

500k+

Records Processed

99.9%

Data Accuracy

States Covered

Architected a Selenium-based extraction system navigating multi-layer dynamically rendered portals across all 36 Nigerian states — standard scrapers fail at portal depth 3+ due to session tokens.
Processed 500k+ records in a 4-week sprint, standardizing outputs to Parquet and loading into a queryable PostgreSQL database (99.9% accuracy vs INEC official source).
Benchmarked ViT-B/16 vs DINOv2-B on brain tumor MRI data with Grad-CAM — DINOv2 outperformed (93.8% vs 91.4%) with better spatial feature localization on tumor boundaries.

SeleniumPostgreSQLPythonPyTorchDINOv2Pandas

Internship

IT Summer Support Intern

Tertiary Education Trust Fund (TETFund) · Abuja

Mar – Sep 2024

90%

Time Reduction

~40 hrs

Saved per Cycle

Report Layouts

Built Python automation pipelines that replaced a 2-day manual Excel report cycle with a batch run under 4 hours — handling 12 different layouts across universities, polytechnics, and colleges of education.
Automated correction logging: every changed cell written to an audit trail so staff could verify and reproduce all transformations without developer involvement.

PythonPandasopenpyxlPostgreSQL

Education

B.Sc. Computer Science

Landmark University · Omu-Aran

2021 – 2025

Algorithms, data structures, machine learning, and software engineering.
Final year thesis on multimodal AI for medical diagnostics — 87.1% ROC-AUC via MedCLIP + TabNet early fusion architecture.

Computer ScienceMachine LearningSoftware Engineering

Project Repository

All Projects

Expand any card for full implementation details, metrics, and what didn't work.

Brain Tumor Classification

PyTorch • ViT • DINOv2 • Grad-CAM

Architecture benchmark on MRI data: Vision Transformer vs DINOv2 with Grad-CAM explainability. Built during Kontemporary internship.

ViT: 91.4%DINOv2: 93.8%Grad-CAMMRI

DINOv2-B: 93.8% accuracy — faster convergence, better tumor-boundary localization on Grad-CAM. Winner.
ViT-B/16: 91.4% accuracy — more epochs required, diffuse attention patterns on scan.
Inference: DINOv2 ~12ms/image vs ViT ~18ms/image on CPU — relevant for clinical deployment.

Key insight

ViT's fixed-grid patch encoding conflicted with MRI's variable-scale features. DINOv2's self-supervised pretraining transferred better despite never seeing MRI data.

Repository private — organizational work. Details available on request.

Human Activity Recognition

PyTorch • Transformers • HAR-70

Real-time activity recognition via Transformer on wearable sensor time-series. 12 classes, 70 participants.

92.3% Accuracy12 Classes70 Participants

92.3% test accuracy across 12 activity classes.
Multi-head self-attention over 6-axis IMU windows — 128-sample windows at 50Hz.
Confusion concentrated on stair-up vs stair-down and standing vs slow-walking — physically ambiguous motion profiles.

Repository

Heart Disease Prediction

Scikit-Learn • XGBoost • SMOTE

Clinical decision support optimized for Recall — minimizing false negatives is the right objective in medical screening.

0.91 Recall0.87 F1Class Imbalance

0.91 Recall on the positive class — optimized to minimize false negatives, the more dangerous error in screening.
XGBoost beat Random Forest (F1 0.87 vs 0.83) after 5-fold stratified CV.
Compared class_weight='balanced' vs SMOTE — SMOTE yielded +0.03 Recall at +0.02 FPR; documented both with trade-off rationale.

Repository

Tech Salary Prediction

Python • XGBoost • Plotly

Regression model predicting tech salaries in USD. Experience outranks location as #1 compensation driver.

R² 0.81MAE ~$9,400Feature Engineering

R² = 0.81, MAE ≈ $9,400 — model explains 81% of salary variance.
Top predictors: experience (34%), remote ratio (21%), company size (17%) — location ranked 4th, not 1st as commonly assumed.
Collapsed 200+ job titles into 12 role families, improving MAE by ~$2,100 vs raw encoding.

Repository

Nigeria Polling Unit Scraper

Python • Selenium • PostgreSQL • Click CLI

Large-scale electoral data infrastructure across 36 states. Multi-layer portal navigation, 500k+ records.

500k+ Records99.9% Accuracy36 States

System Architecture

CLI Entry (Click)
  └── StateWorker(state_id)
        ├── navigate_portal(driver, depth=3)   # session token handling
        ├── retry_handler(max=5, backoff=exp)  # portal uptime
        ├── extract_polling_units(driver)
        └── validate_and_clean(rows)
                  │
          bulk_insert(conn, rows)
                  │
      PostgreSQL: polling_units
      ┌─────────────────────────────┐
      │ state | lga | ward          │
      │ pu_code | pu_name           │
      │ lat | lng | reg_voters      │
      └─────────────────────────────┘

Standard scrapers fail at depth 3+ due to session tokens — solved with explicit wait chains and session state management.
CLI with --resume flag for mid-run recovery without re-scraping completed states.
Outputs to Parquet for analysis + PostgreSQL queryable by state, LGA, and ward.

Repository private — organizational data. Architecture and sanitized schema available on request.

TETFund Report Automation

Python • Pandas • openpyxl • PostgreSQL

Pipeline replacing manual report processing at a federal agency — 2-day cycle cut to under 4 hours.

90% Time Saved~40 hrs/cycle12 Layouts

Pipeline Architecture

Input: /reports/*.xlsx  (12 heterogeneous layouts)
  │
ingest_report(path)
  ├── detect_schema(df)          # 12 layouts + fallback
  ├── clean_missing_values(df)   # forward-fill + flag
  ├── fix_encoding_errors(df)    # cp1252 / utf-8 mixed
  └── validate_funding_cols(df)
  │
export_to_postgres()
  └── upsert on (institution_id, report_period)

Output: structured DB + full audit log

12 different Excel layouts across institution types — each with different column names for the same underlying data.
Every changed cell written to an audit trail so staff could verify all modifications.

Repository private — institutional data. Pipeline design available on request.

Credit Card Fraud Detection

Scikit-Learn • Random Forest • EWMA

Recall-optimized classifier on highly imbalanced financial data. Behavioral drift captured via 7-day sliding-window EWMA features.

Recall-OptimizedEWMA Features0.17% Positive Rate

EWMA of transaction frequency over a 7-day window captures behavioral drift before fraud events.
Optimal threshold selection reduced false negatives by 18% vs default 0.5 at controlled FPR.
SMOTE + cost-sensitive learning for 0.17% positive class rate.

Repository

Other Work

Bunmi Adelugba & Co. — Corporate Website

HTML • CSS • JavaScript • Google Apps Script

Multi-page corporate site for a Chartered Accounting firm. Responsive layout, service pages, contact form.

ResponsiveApps Script

Service pages: Audit, Taxation, Advisory. Fully responsive.
Debugged Google Apps Script form — fixed multipart/form-data vs application/x-www-form-urlencoded mismatch causing silent POST failures.

All Public Repositories