Bioinformatics · Data Science · Scientific AI Evaluation · Pipelines
Salvatore Barbagallo
Combining 7+ years in clinical genomics with computational engineering, building reproducible NGS pipelines, ML models, scientific AI evaluation tasks, and data workflows that transform biological data into decisions.
I combine clinical and genomics laboratory experience with data analysis and process automation to build reliable pipelines, reporting tools, and QC workflows that turn complex biological data into usable results.
My background spans regulated clinical environments where I built Python-driven tools, automated QC pipelines, and worked with NGS from library prep through to clinical reporting.
Through Project Blackwell, I also design deterministic Python-based scientific reasoning tasks that test whether advanced AI systems can infer domain-grounded function behaviour through limited experimental probing rather than simple retrieval.
Currently pursuing an MSc in Bioinformatics at Atlantic Technological University, strengthening computational analysis and reproducible pipeline development and automation.
RegulonML - Real MPRA Regulatory Variant-Effect Modelling
Built a real-data MPRA modelling workflow using public saturation-mutagenesis reporter-assay data to prepare regulatory variant features, train interpretable regression models, and evaluate variant-effect prediction with leakage-aware element-level splitting.
View details
Problem: Regulatory variant-effect modelling requires careful handling of MPRA count data, assay metadata, and train/test leakage between variants from the same regulatory element.
Action: Added scripts to download public Kircher saturation-mutagenesis MPRA data, filter promoter elements, engineer variant/count features, and train Elastic Net and Random Forest regression models.
Impact: Created a defensible real-data regulatory-genomics workflow with documented provenance, element-level group splitting, permutation importance, and clear limitations around full promoter-design claims.
Built an end-to-end e-commerce data engineering project covering raw data ingestion, SQL/NoSQL querying, dimensional warehouse design, ETL automation, PostgreSQL analytics, BI reporting, and Spark workflow practice.
View details
Objective: Build a structured e-commerce data platform that connects operational sales data, NoSQL product data, warehouse modelling, reporting outputs, and scalable analytics practice.
Approach: Reorganised the IBM Data Engineering capstone into a clean portfolio repository, fixed warehouse SQL inconsistencies, loaded dimensional tables into PostgreSQL, and generated validated analytical CSV outputs.
Outcome: Produced a reproducible data engineering artefact with 300,000 fact rows loaded, warehouse validation notes, sales aggregations by country/category, rollup and cube outputs, ETL scripts, MongoDB commands, Spark notebook, and BI report.
SoftCart data platform architecture showing the flow from customer-facing web activity into MySQL and MongoDB, through PostgreSQL staging and DB2 warehousing, with Cognos Analytics for BI reporting and Spark for large-scale analytics.
Data EngineeringPostgreSQLSQLMongoDBETLData WarehouseSparkBI Reporting
M. tuberculosis WGS Variant Analysis Workflow (Galaxy)
Built an end-to-end WGS workflow for Mycobacterium tuberculosis drug resistance analysis, standardising variant detection and interpretation across QC, alignment, annotation, and read-level validation.
View details
Problem: Resistance interpretation from WGS data requires a long multi-step workflow with multiple failure points and high risk of inconsistent analysis.
Action: Built and executed a Galaxy-based workflow covering QC, trimming, alignment, coverage analysis, variant calling, annotation, and IGV-based validation.
Impact: Produced a structured and reproducible resistance analysis workflow supporting clearer interpretation of clinically relevant loci.
Framed Galaxy workflow for M. tuberculosis WGS variant analysis
Eliminated inconsistent and error-prone RNA-seq analyses by building a containerised workflow that standardises processing from raw FASTQ to differential expression outputs.
View details
Problem: RNA-seq analyses are often manually assembled, environment-dependent, and difficult to reproduce across machines.
Action: Built a containerised Nextflow pipeline integrating FastQC, Cutadapt, STAR, featureCounts, MultiQC, and DESeq2 into a one-command workflow.
Impact: Reduced setup friction, improved reproducibility, and produced standardised outputs suitable for scalable downstream expression analysis.
Proof run: Nextflow test execution (Docker profile)
Deep Learning Experiments (Keras vs PyTorch, CNN, ViT)
Comparative deep learning experiments using Keras and PyTorch across CNNs, Vision Transformers, and transfer learning for image classification tasks.
View details
Objective: Compare deep learning approaches and frameworks across multiple computer vision tasks.
Approach: Implemented CNNs, transfer learning models, and Vision Transformers in both Keras and PyTorch, evaluating performance and training behaviour.
Outcome: Demonstrated differences in performance, flexibility, and workflow between frameworks and architectures.
Deep LearningPyTorchKerasCNNVision TransformerTransfer LearningComputer Vision
Segmented 7,043 telecom customers using K-Means, hierarchical clustering, and DBSCAN, identifying a high-risk segment with ~42% churn compared to ~14% in the most stable group.
View details
Objective: Identify meaningful customer segments to enable targeted retention strategies instead of treating churn as a uniform problem.
Approach: Applied clustering methods on standardised features and selected the final model using silhouette score.
Outcome: Produced interpretable segments that can guide retention campaigns, onboarding improvements, and pricing strategies.
PCA projection of customer clusters (K-Means, k=3)
Transformed fragmented wearable data into a structured analytics dataset to uncover behavioural patterns across activity, sleep, heart rate, and weight for 35 users.
View details
Problem: Raw Fitbit data was split across multiple files and not directly usable for behavioural analysis or product-facing insight generation.
Action: Cleaned, merged, and summarised wearable datasets using R, dplyr, and SQLite, then created a Tableau-ready user summary with 15 behavioural and wellness variables.
Impact: Enabled clearer identification of activity, sleep, and health trends to support data-driven product and marketing recommendations.
Google Fiber - Contact Centre Analytics & BI Dashboard
Identified repeat-contact patterns across 85,179 customer interactions by consolidating fragmented regional datasets into a unified analytics view, enabling clearer visibility into service inefficiencies.
View details
Problem: Customer contact data was split across regional datasets, making it difficult to understand repeat-contact behaviour and pinpoint service pain points.
Action: Consolidated 3 datasets into a unified analysis table covering 1,350 records and 85,179 contact events across 5 contact types, then built a Tableau dashboard to surface repeat-contact patterns and operational friction.
Impact: Turned fragmented customer-support data into a decision-ready BI view that can support targeted service improvements and reduction of avoidable support load.
Built predictive models on 14,999 employee records to identify the drivers of 23.8% turnover and translate model outputs into retention-focused recommendations.
View details
Problem: The business faced substantial employee turnover, but the main drivers of attrition were not clearly understood.
Action: Performed exploratory analysis, engineered features, and trained classification models using Python, scikit-learn, and XGBoost to analyse attrition patterns across 10 departments.
Impact: Produced evidence-based recommendations to support workforce retention strategy and prioritise the factors most associated with employee loss.
Data ScienceMachine LearningPythonscikit-learnXGBoostHR Analytics
Analysed 19,382 posts to uncover highly skewed engagement patterns and identify content features associated with stronger performance.
View details
Problem: Engagement was unevenly distributed, making it difficult to understand which content characteristics were linked to stronger performance.
Action: Used Python, pandas, and visual analytics to explore engagement metrics, inspect distribution patterns, and examine relationships between content features and post performance.
Impact: Generated evidence-based recommendations to support content optimisation and stronger audience engagement strategy.
Data AnalyticsExploratory Data AnalysisPythonPandasMatplotlibPlotly
High-level AWS architecture design for migrating on-premises workloads to a cloud-native, fully managed solution, ensuring scalability, fault-tolerance, and operational efficiency.
View details
Objective: Migrate two on-prem workloads, a three-tier web application and a Hadoop-based analytics environment, into a modern AWS environment with managed services.
Approach: Designed an end-to-end cloud solution using AWS managed services including CloudFront, S3, ECS on Fargate, ALB, Aurora MySQL, ElastiCache, SQS, EMR, Glue, Athena, Redshift, and QuickSight.
Outcome: Produced a decoupled, fault-tolerant, multi-AZ architecture that modernises both workloads while reducing operational overhead through managed services.
AI Training & Evaluation WorkMercor · Micro1 · OutlierOpen
Mercor
Project Blackwell
Computational STEM Masters – Bioinformatician
Designing frontier scientific-reasoning evaluations that test whether advanced language models can infer hidden biological and computational rules through controlled experimentation.
Build deterministic Python black-box functions, discriminative unit tests, and tightly constrained scientific prompts.
Create biology-grounded benchmark problems requiring hypothesis generation, abstraction, and first-principles reasoning rather than factual recall.
Validate task behaviour and failure modes to ensure reliable, reproducible model evaluation.
Delivered processing and cryopreservation of PBSCs, bone marrow, DLI, and CD34+ enriched products in a regulated clinical environment.
Generated flow cytometry data on CD3+ and CD34+ populations for time-sensitive clinical decision-making.
Supported 11 active clinical and ATMP trials, including CAR-T therapies, with responsibility for traceability, documentation quality, and audit-ready data handling in a high-stakes regulated clinical setting.
Designed Python/Excel tracking tools replacing manual reconciliation, reduced inventory errors by ~30%, saved ~10 hours/week.
Led digitisation of SOPs and QA documentation, standardising data handling across workflows.
CooperGenomics
Jul 2019 – Sep 2021
London, UK
Laboratory Scientist · Clinical Genomics
Processed embryo samples for PGT-A, PGT-SR, and PGT-M testing within a high-throughput clinical genomics pipeline.
Delivered NGS library preparation and QC across 96–192 samples/run; produced 500+ clinical reports per week.
Programmed and validated Mosquito HV and Dragonfly liquid handlers, improving workflow scalability and reproducibility.
Contributed to SOP writing and review, strengthening ISO-compliant laboratory practice.
Leicester Royal Infirmary
Dec 2018 – Jun 2019
Leicester, UK
Biomedical Laboratory Assistant · Cytology
Managed sample reception and prepared specimens for Papanicolaou staining.
Maintained reagents and ensured sample integrity end-to-end.
SAS Programming 1: Essentials · SAS Programming 2: Data Manipulation Techniques
Wellcome
Bioinformatics for Biologists: Linux, Bash, R · Analysing Genomics Datasets
Coursera
Access Bioinformatics Databases with Biopython
freeCodeCamp
Data Analysis with Python · Relational Databases · Scientific Computing
Le Wagon
Data Visualization with Tableau
DE<code>LIFE
Genomes, Networks & Pathways · Data Science & Machine Learning
Contact
Let's Connect
Open to opportunities in bioinformatics, scientific AI evaluation, data science, and data pipelines and automation. Based in London and remote-friendly.
