Dr Delphine Vincent | Data Science, Machine Learning, Natural Language Processing & Scientific Analytics

Biography

Research Scientist (1998-2023)

I have accrued 25 years of experience in life science research, quantitative analysis, bioinformatics, and scientific communication.

Academic and Research Career

Ph.D. in Plant Sciences awarded with Highest Distinction in 2003.
Post-doctoral research conducted in the USA, France, and Australia (2003–2013).
Senior Research Scientist for the Victorian State Government (2013–2023).

Quantitative and Statistical Expertise

Advanced use of Excel, SQL, R, and Python for handling, transforming, cleaning, and integrating qualitative and quantitative datasets.
Expertise in descriptive, univariate, and multivariate statistical analyses, including regression, ANOVA, PCA, SOM, HCA, k-means clustering, and heat maps.
Experience using multiple analytical environments including Python, R, Genedata Expressionist, SAS, JMP, and Statistica.
Strong experience in translating complex datasets into interpretable scientific conclusions.

Bioinformatics and Scientific Data Interpretation

Extensive use of bioinformatics platforms and biological databases for sequence analysis, functional annotation, pathway reconstruction, protein structure interpretation, and multi-omics integration.
Regular use of tools such as BLAST, UniProt, KEGG, STRING, Cytoscape, TAIR, Reactome, BioCyc, MetaboAnalyst, ShinyGO, REVIGO, AlphaFold, and proteomics repositories.
Experience integrating transcriptomics, proteomics, metabolomics, genome annotation, and public biological knowledgebases to support biological interpretation.
Strong capacity to move from raw biological outputs to mechanistic hypotheses through combined computational and biological reasoning.

bioinformatics tools overview

Scientific Communication and Leadership

Extensive data visualisation experience using Python, R, Power BI, Tableau, and Excel across diverse chart types and storytelling formats.
Strong expertise in scientific storytelling, including schematic diagrams and publication-ready visual summaries.
Codes and datasets made publicly available through GitHub, Tableau Public, MassIVE, ProteomeXchange, and Apollo JBrowse.
Reviewer and editor for multiple scientific journals over the past 22 years.

sankey chart of research themes and species

Data Scientist (2024-now)

I have now become a data science, machine learning (ML), and Natural Language Processing (NLP) specialist, as well as a front-end developer, by undertaking the following certifications and applied projects:

1/ CodeCademy Career Path: Data Scientist - Analytics Specialist

Training focus: SQL, Python 3, Tableau, Excel, and GitHub (syllabus).

Applied outcome: I applied my data science and Python skills to the largest proteogenomics dataset to date to refine wheat genome annotation, leading to a published study in large-scale wheat proteogenomics (DOI 10.3390/ijms25168614).

Further details: More details are available in "Big Data".

Python generated charts

2/ CodeCademy Career Path: Data Scientist - Machine Learning

Training focus: Supervised and unsupervised learning, ensemble methods, Support Vector Machines, recommender systems, Naive Bayes, deep learning, and neural networks using Python (syllabus).

Applied outcome: I subsequently applied these skills across a diverse portfolio of research and analytics projects spanning multi-omics, metabolomics, bibliometrics, anomaly detection, recommendation systems, and scientific career mapping. These projects combined exploratory data analysis (EDA), feature engineering, clustering, supervised and unsupervised machine learning, natural language processing (NLP), and interactive visualisation to uncover interpretable patterns in complex biological, behavioural, and textual datasets.

Further details: For a full list of my ML and NLP projects, see "Machine Learning and NLP".

Python generated charts

3/ CodeCademy Skill Path: Build a Website with HTML, CSS and GitHub Pages

Training focus: Website structure, styling, deployment, and version control using HTML, CSS, JavaScript, and GitHub Pages (syllabus).

Applied outcome: Using HTML, JavaScript, and CSS, I created my own professional website — the very one you are currently browsing. I also designed and deployed a website for Data Biome.

Further details: More information on my website workflow is available in "CodeCademy Skill Path - Build a website" and the Data Biome website project.

website screenshot

4/ CodeCademy Course: Learn Prompt Engineering

Training focus: Prompt design strategies for productive and reliable interaction with generative AI systems (syllabus).

Applied outcome: This course strengthened my ability to interact efficiently with generative AI systems and to integrate them thoughtfully into scientific, analytical, and creative workflows.

Further details: More details are available in "CodeCademy Course - Learn Prompt Engineering".

prompt engineering illustration

5/ CodeCademy Skill Path: Build a Chatbot with Python

Training focus: Python-based chatbot design and implementation (syllabus).

Applied outcome: I learned how to create a chatbot with Python and implemented a closed-domain chatbot for this very website. You can find it in the bottom right corner of the site. Give it a try — responses may take a little time to appear.

Further details: More details on the development of my chatbot are available in "CodeCademy Skill Path - Build a chatbot in Python".

chatbot illustration

6/ CodeCademy Course: Large Language Model (LLM)

Training focus: Foundations of large language models and text-based generative artificial intelligence (syllabus).

Applied outcome: This course gave me a solid grounding in how LLMs work and how they can support text analysis, content generation, retrieval, and human–AI interaction.

LLM illustration

7/ Automating Visual Storytelling with Python

Creative workflow: Using Python and FFmpeg, I automate the creation of aesthetic photo montages with ambient sounds from my large collection of digital images.

Applied outcome: These automated visual stories are published on YouTube and TikTok as part of my creative project.

Further details: More details are available here.

YouTube banner

Work with Me

I am available for consulting projects in data science, machine learning, NLP, bioinformatics, and scientific analytics.

Independent Consultant

I operate as an independent consultant in Australia under a registered sole-trader business structure.

ABN: 74 267 801 710
Consulting focus: data science, machine learning, NLP, bioinformatics, scientific analytics, and technical reporting
Working style: remote-friendly, project-based, transparent, and reproducible

How I Can Help

clean, structure, and integrate complex datasets
design statistical, machine learning, or NLP workflows tailored to your research or business question
perform exploratory data analysis, feature engineering, modelling, and interpretation
support biomarker discovery, classification, prediction, clustering, text mining, and semantic analysis
develop reproducible Python notebooks, Tableau or Power BI dashboards, reports, and presentation-ready figures
translate complex results into clear, decision-oriented insights for technical and non-technical stakeholders

Typical Deliverables

cleaned and analysis-ready datasets
reproducible Python or R workflows
statistical and machine learning reports
publication-ready charts and summary figures
interactive dashboards or prototype interfaces where relevant
technical slide decks, scientific write-ups, and project recommendations

Suitable Projects

omics, proteomics, metabolomics, transcriptomics, proteogenomics and bioinformatics projects
scientific or technical datasets requiring advanced analytics
NLP projects involving free text, topic modelling, semantic similarity, or document analysis
custom machine learning workflows for classification, prioritisation, anomaly detection, or prediction
research translation, reporting, and data storytelling for grants, publications, or stakeholders

Contact Form

This form is intended for genuine consulting, collaboration, or speaking enquiries.

Please provide a short project summary using your professional or institutional email address where possible.

Full name *

Organisation / company

Work email *

Project type *

Preferred timeline

Indicative budget

Project summary *

Skills

Statistical Analyses, ML and NLP

I perform a wide range of statistical analyses using industry-standard tools such as Genedata Expressionist Analyst, SAS, R, JMP, Statistica, and Python, delivering rigorous, reproducible, and interpretable insights across biological, experimental, and complex multidimensional datasets.

descriptive statistics & exploratory data analysis

summary statistics (means, medians, variances, quartiles, standard deviations, standard errors)
distribution analysis (histograms, density plots, QQ plots, skewness inspection)
data quality checks, outlier detection, and missing value exploration
visual exploration using box plots, scatterplots, violin plots, heat maps, volcano plots, treemaps, and correlation matrices
identification of hidden patterns, trends, and anomalies prior to formal modelling

univariate inference & statistical testing

correlation analysis and association testing
linear regression and general trend modelling
analysis of variance (ANOVA)
t-tests and non-parametric alternatives when assumptions are not met
effect size estimation and multiple-testing aware interpretation

multivariate analysis & structure discovery

principal component analysis (PCA)
partial least squares (PLS)
hierarchical clustering analysis (HCA)
k-means clustering
self-organising maps (SOM)
UMAP and other dimensionality reduction methods for pattern discovery and visual interpretation

data preprocessing, normalisation & quality control

normalisation and scaling (log, Pareto, Z-score, quantile, Yeo–Johnson)
batch effect detection and correction
QC-based filtering, drift monitoring, and noise reduction
feature selection, variance filtering, and transformation of skewed variables
preparation of robust, modelling-ready datasets from noisy experimental or observational data

I design and implement a wide range of Machine Learning (ML) and Natural Language Processing (NLP) workflows using Python within a reproducible Conda and Jupyter Notebook environment. My work spans supervised, unsupervised, and deep learning approaches, with strong emphasis on feature engineering, interpretability, validation, and deployment-oriented outputs.

supervised machine learning

linear and logistic regression
regularised models (Lasso, Ridge, Elastic Net)
support vector machines (SVM)
decision trees and random forests
gradient boosting methods (XGBoost, LightGBM, CatBoost)
k-nearest neighbours (k-NN)
naïve Bayes classifiers
ensemble and stacked models for improved predictive performance
classification, regression, ranking, prioritisation, and predictive scoring

unsupervised machine learning

k-means clustering
hierarchical clustering analysis (HCA)
self-organising maps (SOM)
density-based clustering (DBSCAN, HDBSCAN)
dimensionality reduction (PCA, UMAP, t-SNE)
topic modelling and latent structure detection (LDA, NMF, BERTopic)
similarity-based matching, grouping, and anomaly detection

deep learning, embeddings & NLP

multi-layer perceptron (MLP) neural networks
recurrent neural networks (LSTM, GRU)
transformer-based models (BERT, DistilBERT)
tokenisation, lemmatisation, TF-IDF, and feature extraction from free text
sentence embeddings, semantic similarity, and document clustering
text classification, topic extraction, corpus structuring, and keyword analysis
integration of textual and structured features into unified analytical pipelines

model evaluation, interpretation & delivery

cross-validation, threshold tuning, and hyperparameter optimisation
evaluation using accuracy, precision, recall, F1-score, AUC, RMSE, MAE, R², silhouette, and related metrics
feature importance analysis, coefficient interpretation, and explainable outputs
delivery of reproducible code, publication-ready figures, dashboards, reports, and interactive interfaces
deployment-oriented thinking from research question to actionable output

Data Visualisation

Throughout my career, I've created many visualisations to describe categorical and quantitative data:

distribution

one variable:

histogram
box and wisker plot
density plot
violin plot

two variables:

scatterplot
stacked histogram
pyramid plot

text data:

wordcloud

relationship

two variables:

scatterplot
heatmap
volcano plot

three or more variables:

scatter bubble plot
chord diagram
network diagram
pair grid plot

comparison

static

pie chart
donut chart
waterfall chart
stacked bar chart
tree map
Venn diagram

change over time

stacked bar chart
stacked area chart
100% stacked bar/area chart

composition

among items

few categories

vertical bar chart
horizontal bar chart

many categories

pair plots
heatmap

over time

cyclical data

circular area chart

non cyclical data

line chart
Gantt chart

few categories

vertical bar chart

many categories

line chart

geographic

map using latitude and longitude

Portfolio

Big Data Projects on Wheat

Wheat (Triticum aestivum) is an important crop whose reference genome offers a valuable resource for understanding wheat genetic structure, improving agronomic traits, and developing new cultivars.
I was tasked to develop a processing and analytical workflow to screen the proteome (all the proteins) of thousands of cultivars.

I first developped a fast, accurate and reproducible method which I then applied to 4,087 wheat grain samples.
I produced, managed and analysed close to 8,000 files representing >1.2 Gb of data.

I employed sophisticated methods of statistical analyses, bioinformatics amd data mining to generate meaningful biological insights.

Proteins were then mapped onto the wheat genome (version 2) which not only validated genes of high confidence, but also promoted genes of low confidence as well as detect novel genes.

Links to the scientific publications are provided below.

Optimising the protein extraction and digestion from wheat grains
>> Mining the Wheat Grain Proteome

Screening the proteome of >4,000 wheat cultivars
>> A community resource to mass explore the wheat grain proteome and its application to the late-maturity alpha-amylase (LMA) problem

Mapping the wheat proteins onto the genome
>> Community Resource: Large-Scale Proteogenomics to Refine Wheat Genome Annotations

I have recently revisited the dataset listing the peptide profiles of >4,000 wheat cultivars in order to classify them using machine learning algorithms, and determine peptides contributing most to these classes (article in preparation).
More details can be found at "Machine Learning and NLP".

Machine Learning and NLP

This section showcases selected machine learning and natural language processing projects spanning multi-omics, metabolomics, bibliometrics, anomaly detection, recommendation systems, and scientific career analytics. Each project combines reproducible Python workflows with interpretable visual outputs and, where relevant, publication-oriented reporting.

FNRL Mode-of-Action in Arabidopsis Roots: an Integrated Multi-Omics, Machine Learning and Bioinformatics Study

Objective: Investigate the function of the root-specific FNRL gene in Arabidopsis thaliana by integrating phenotyping with transcriptomics, proteomics, metabolomics, machine learning, and bioinformatics. The study compared wild type plants, two independent loss-of-function mutants, and a GFP-complemented line to uncover molecular pathways associated with FNRL regulation.

Dataset: Multi-omics and phenotypic dataset comprising root RNA-seq, LC-MS/MS proteomics, GC-MS metabolomics, nitrate content measurements, and root length data across four genotypes and biological replicates.

Methods: Data cleaning and harmonisation, missing-value imputation, log transformation, cross-omics scaling, PCA, differential expression analysis, biomarker discovery based on mutant consistency and GFP rescue rules, hierarchical clustering, k-means clustering, LDA, Elastic Net and Random Forest modelling, plus bioinformatics mining using TAIR, UniProtKB, GO/AmiGO, KEGG, STRING, Cytoscape, PaintOmics and AraCyc.

Identified 504 high-confidence FNRL-regulated biomarkers across transcriptomics, proteomics, and metabolomics.
Resolved biomarker behaviour into 4 main expression profiles, separating FNRL-activated and FNRL-repressed targets with strong or partial genetic rescue.
Built a conceptual model of FNRL regulation based on two main axes: regulatory direction and rescue strength.
Showed that root growth phenotypes were highly predictable from biomarker-derived molecular profiles, with Elastic Net models achieving strong performance.
Integrated pathway and interaction analyses highlighted processes linked to nitrogen metabolism, protein processing, ubiquitin-mediated proteolysis, and root-associated regulation.
Generated an interpretable multi-omics resource to support biological interpretation of FNRL mode of action and downstream manuscript preparation.

Technical highlights: Python • multi-omics integration • biomarker discovery • clustering • LDA • Elastic Net • Random Forest • pathway analysis • network biology

Status: Article in preparation

Final report Python code and Jupyter notebook FNRL study overview

Machine Learning Reveals Synergistic Effects of Lactobacillus helveticus in Camel Milk Fermentation Using Metabolomics

Objective: Develop a computational biomarker discovery pipeline to isolate metabolites specifically associated with Lactobacillus helveticus fermentation in camel milk, and determine whether co-culture with L. bulgaricus or S. thermophilus induces synergistic or antagonistic metabolic effects.

Dataset: Untargeted UPLC-MS/MS metabolomics dataset covering 13,400 metabolites, including 3,632 identified compounds, generated from camel and bovine milk fermented under mono- and co-culture conditions.

Methods: Three-way ANOVA, PCA, LDA, HDBSCAN, k-means, spectral clustering, self-organising maps (SOM), Random Forest classification, feature importance ranking, metabolite annotation, superclass analysis, and pathway exploration using MetaboAnalyst and external metabolic databases.

Designed a 4-level biomarker selection pipeline reducing 13,400 compounds to the most biologically relevant features.
Selected 2,069 metabolites through full-factor interaction testing, then refined to 1,017 metabolites showing clear synergy/antagonism patterns.
Applied Random Forest modelling to isolate 508 high-impact biomarkers associated with metabolic interactions driven by L. helveticus.
Retrieved 133 identified metabolites, of which 87 displayed synergistic profiles and 46 antagonistic profiles.
Highlighted metabolite classes linked to amino acid metabolism, bioactive lipids, carbohydrates, and fermentation-associated functional compounds.
Developed a reusable in silico protein digestion tool to simulate enzyme cleavage of camel and bovine caseins, supporting interpretation of dairy proteolysis and related cheese studies.
Delivered a robust analytical framework supporting downstream biological interpretation and manuscript preparation.

Technical highlights: Python • metabolomics • feature selection • clustering • Random Forest • biomarker discovery • pathway analysis • Side tool: in silico protein digestion

Status: Article in preparation

Final report Python code and Jupyter notebook Related bioinformatics tool Associated publication camel milk study overview

Machine Learning Modelling of Synergy and Antagonism of Phenolic Compounds Using Antioxidant Assays

Objective: Develop a machine learning framework to quantify synergistic and antagonistic interactions between phenolic compounds across multiple antioxidant assays, using observed-versus-predicted absorbance behaviour to reveal non-additive biochemical effects.

Dataset: Experimental antioxidant dataset comprising individual standards and binary mixtures (CB1–CB5) tested across multiple assay systems, with absorbance measurements collected for pure compounds and combinations under controlled concentrations.

Methods: Exploratory data analysis, assay-wise regression modelling, XGBoost prediction, residual-based synergy scoring, threshold optimisation, confusion matrix evaluation, feature importance ranking, and comparative interpretation across antioxidant assay types.

Built XGBoost regression models to predict expected absorbance values from single-compound behaviour.
Calculated synergy and antagonism scores from prediction residuals, enabling quantitative classification of interaction strength.
Demonstrated that interaction behaviour varies substantially across assay systems, revealing assay-dependent biochemical responses.
Generated interpretable classification outputs distinguishing synergistic, additive, and antagonistic mixtures.
Established a transferable computational framework for analysing compound interactions in antioxidant chemistry.
Produced publication-ready visual outputs supporting article submission.

Technical highlights: Python • XGBoost • regression modelling • residual scoring • assay comparison • feature interpretation

Status: Article under review

Final report Python code and Jupyter notebook antioxidant assay study overview

Bread Protein Biomarker Discovery Assisted by Machine Learning

Objective: Identify peptide biomarkers associated with wheat genotype groups and flour-quality protein expression, using statistical learning and machine learning to support biomarker-assisted selection in bread wheat.

Dataset: Large-scale proteomics dataset comprising thousands of peptides derived from flour proteins quantified across a broad panel of wheat genotypes, integrated with genotype metadata and protein annotation.

Methods: Data normalisation, correlation analysis, t-tests, ANOVA, hierarchical clustering, k-means clustering, self-organising maps (SOM), PCA, LDA, Random Forest, SVM, MLP neural network, and stacked ensemble modelling for genotype classification and biomarker ranking.

Integrated statistical and machine learning outputs into a unified biomarker discovery framework.
Resolved genotype structure through multiple complementary clustering approaches, including HCA, k-means, and SOM.
Applied supervised models to classify genotype groups using peptide abundance profiles.
Built a stacked ensemble model combining Random Forest, SVM, and MLP to improve classification robustness.
Identified peptide biomarkers linked to key flour-quality protein groups, including glutenins and gliadins.
Generated biologically interpretable candidate markers relevant for wheat breeding and flour functionality.

Technical highlights: Python • proteomics • biomarker discovery • clustering • stacked machine learning • classification • feature ranking

Status: Article in preparation

Final report Python code and Jupyter notebook wheat proteomics biomarker charts

Career Trajectory Mapping From My Scientific Publications Using NLP, Machine Learning and Analytics

Objective: Develop a reproducible NLP and machine learning framework to analyse the thematic evolution of a scientific career through publication metadata, abstracts, keywords, and semantic content, with the goal of identifying major research phases, transitions, and future directions.

Dataset: Curated corpus of personal scientific publications spanning multiple years, integrating titles, abstracts, keywords, journal metadata, authorship patterns, and publication timelines.

Methods: Text preprocessing, tokenisation, TF-IDF, topic modelling, keyword frequency analysis, semantic clustering, temporal trend analysis, PCA, LDA, and machine learning-assisted interpretation of research trajectory patterns.

Built a structured NLP corpus from publication records across multiple scientific domains.
Identified major thematic transitions across career stages using topic modelling and semantic clustering.
Mapped the evolution from molecular plant science to multi-omics, machine learning, and computational biology.
Integrated publication chronology with thematic outputs to reconstruct career progression pathways.
Generated visual analytics highlighting dominant research themes, emerging directions, and interdisciplinary expansion.
Produced a data-driven framework transferable to researcher profiling, strategic planning, and academic portfolio analysis.

Technical highlights: Python • NLP • topic modelling • TF-IDF • semantic analysis • temporal analytics • machine learning interpretation

Status: Completed analytical report

Final report Python code and Jupyter notebook career trajectory NLP charts

Aliens, Algorithms & Anomalies: Visualising the NUFORC UFO Sightings

Objective: Develop a machine learning and NLP framework to analyse large-scale UFO sighting reports, identify reporting patterns, classify anomalous events, and prioritise cases with high investigative value.

Dataset: NUFORC database containing more than 150,000 UFO sighting reports, integrating temporal records, geographical metadata, witness descriptions, event characteristics, and free-text narratives.

Methods: Data cleaning, geospatial enrichment, NLP feature extraction from witness reports, exploratory data analysis, temporal and spatial visualisation, classification modelling, anomaly prioritisation, and predictive scoring.

Cleaned and enriched large-scale historical sighting data with geographical coordinates and regional metadata.
Extracted structured variables from free-text witness narratives to support NLP-driven analysis.
Built predictive models to identify high-priority sightings associated with stronger anomaly indicators.
Developed case prioritisation logic to flag reports with elevated close-encounter characteristics.
Generated multi-scale visual analytics showing temporal waves, spatial hotspots, and reporting behaviour.
Combined scientific rigour with unconventional public data to demonstrate transferable anomaly-analysis methodology.

Technical highlights: Python • NLP • geospatial analytics • classification • feature engineering • anomaly scoring • Tableau

Status: Completed analytical report

Final report Python code and Jupyter notebook Interactive Tableau dashboard ufo anomaly analytics charts

Decoding Love with Data: Matchmaking in a Dating App

Objective: Develop a machine learning and NLP framework to identify compatible matches between dating profiles by combining structured demographic features, personal preferences, and free-text self-descriptions.

Dataset: Large-scale dating profile dataset including demographic variables, lifestyle attributes, preferences, and millions of words extracted from personal essays written by users.

Methods: Text preprocessing, tokenisation, lemmatisation, topic modelling (LDA), feature encoding, clustering, cosine similarity, nearest-neighbour matching, and interactive match retrieval.

Processed millions of words from user essays to extract meaningful lifestyle and personality themes.
Applied topic modelling to identify dominant discussion themes across personal profiles.
Integrated structured profile variables with NLP-derived features into a unified similarity framework.
Built cluster-based matching logic to improve candidate relevance before similarity scoring.
Computed profile-to-profile similarity using cosine similarity within behavioural clusters.
Developed an interactive interface to retrieve top compatible matches for any selected profile.

Technical highlights: Python • NLP • topic modelling • clustering • cosine similarity • recommendation logic • Gradio interface • Tableau

Status: Completed machine learning project

Final report Python code and Jupyter notebook Interactive Tableau dashboard dating app NLP matching charts

Mapping the Landscape of Scientific Publications

Objective: Explore large-scale scientific publication metadata to identify bibliometric trends, publication patterns, dominant research topics, and long-term shifts in disciplinary focus.

Dataset: Scientific article metadata dataset containing approximately 120,000 publications with information on titles, authors, publication dates, journals, publishers, languages, article types, references, and subject descriptors.

Methods: Metadata cleaning and wrangling, exploratory data analysis, bibliometric visualisation, subject text preprocessing, TF-IDF vectorisation, and k-means clustering to group fine-grained subject labels into broader thematic categories.

Cleaned and structured a large bibliographic dataset to support interpretable publication analytics.
Analysed long-term trends in publication volume, article types, journals, publishers, languages, citations, and title length.
Identified prolific authors and high-output journals across multiple scientific fields.
Explored subject evolution over time to distinguish long-lasting, emerging, and short-lived research areas.
Reduced 1,265 subject labels into 7 broader thematic clusters using text mining and k-means clustering.
Produced publication-ready visual analytics and an interactive Tableau dashboard for bibliometric exploration.

Technical highlights: Python • bibliometrics • metadata analytics • text mining • TF-IDF • k-means clustering • Tableau

Status: Completed analytical report

Final report Python code and Jupyter notebook Interactive Tableau dashboard Raw dataset scientific publication analytics charts

Visual Automation

I am a keen photgrapher and have taken a plethora of digital pictures for the past 3 decades. In May 2025, I started to share them on social media.
My chanels @60_seconds_of_calm showcases my travels, garden, and travels.
I am using a Python program to automate the creation of an aesthetic photo montage using Python and FFmpeg.
It processes a folder of images, applies fade transitions, and enriches the final video with ambient sounds and a piano music.
There are 15 themes (castle, church, national parks, cities, villages, garden, art, whimsical etc...) which are processed in batch.
A stylised "door opening & closing" animation that I creayed overlays the video to simulate stepping outside for a 1 min break.
Each video contains 20 images, which are being processed as described below.
You can view those videos on YouTube and TikTok.

Technologies Used

Python (OpenCV, Pillow)
FFmpeg (Video compression, audio mixing, overlay effects)
Command-line automation using subprocess

Workflow Steps

Load Images: Count all JPG/PNG files in a folder.
Resize & Prepare: Each image is padded and resized to fit a 1920×1080 frame while preserving its aspect ratio.
Montage Creation: A video is generated with each image shown for 2 seconds, and smooth crossfades between them using OpenCV.
Video Compression: The montage is compressed with FFmpeg using the H.264 codec (CRF=28) to reduce file size.
Ambient Sound: A high-quality nature soundscape is layered under the video with fade-in and fade-out effects.
Background Music: A piano loop is blended with the ambient sound using precise volume and fading controls.
Door Animation Overlay: A stylised "door opening" appears at the beginning, and a "door closing" fades in at the end, overlayed on the video.
Final Output: A beautifully compressed MP4 (1080p, H.264) video montage, complete with ambient atmosphere and musical mood.

Website design

I designed and deployed the website for Data Biome ( URL), a biotechnology startup developing probiotics for ruminants to reduce methane emissions and mitigate greenhouse gas impact. This project covered the full pipeline from local development to WebCentral production hosting, including SMTP infrastructure for the contact form.

Project setup & structure
- Created full local project structure with index.html, style.css, reset.css, main.js, and asset directories.
- Versioned ZIP deployments for incremental hosting uploads.
- Defined brand colour palette (Powder Blue #B0E0E6, Dodger Blue #1E90FF, Navy #000080, Dark Orange #F54927).
Information architecture & layout
- Designed full one-page layout: Home, Projects, Science, Partners, People, Contact.
- Built reusable responsive card system (image left, text right).
- Sticky, accessible top navigation with keyboard and focus support.
Branding & visual identity
- Integrated custom logo, icons, and typographic identity.
- Implemented background imagery with contrast overlays for readability.
- WCAG-aware contrast and scalable typography across breakpoints.
Hero background animation (video engineering)
- Full-width fixed MP4 background with layered z-index safety.
- Advanced FFmpeg processing via Python subprocess: slowing, mirroring, flipping, stacking (side-by-side, top–bottom, 2×2 grids).
- Web-optimised H.264 export (-crf, -preset, +faststart).
Accessibility & UX
- Semantic HTML5 landmarks and keyboard-navigable menus.
- Clear focus states and form validation feedback.
- Reduced-motion considerations for animated elements.
Contact form (secure SMTP infrastructure)
- Client-side validation with honeypot bot protection.
- Server-side mail handling via contact.php and PHPMailer.
- SMTP relay configured through WebCentral mail servers.
- DNS authentication repaired (SPF + DMARC).
- End-to-end validation: browser → PHP → SMTP → inbox.
WebCentral-compatible CSS engineering
- Strict compatibility with WebCentral legacy CSS linter.
- Removed unsupported modern functions (clamp, gap, color-mix, logical shorthands).
- Implemented hex fallbacks before RGBA and classic layout properties.
Server setup (cPanel)
- Created mailboxes (info@, no-reply@).
- Configured SMTP authentication and relay permissions.
- Staging environment under /public_html/dev/ for private review.
Deployment & caching strategy
- ZIP deployment via File Manager.
- Cache-buster versioning for CSS/JS files.
- Linux-safe path validation and case-sensitive testing.
QA & cross-environment validation
- Reconciled differences between VS Code Live Server and hosted WebCentral rendering.
- Mobile, tablet, desktop responsive validation.
- Form delivery verified via live inbox testing.
Production launch
- Final deployment to /public_html.
- Live navigation, media, and form tested post-launch.
- Client handover for content updates and safe maintenance.
Ongoing maintenance
- Monthly content updates and link checks.
- Form delivery health monitoring.
- Roadmap for future feature expansion.
Key challenges solved
- Diagnosed and resolved silent SMTP delivery failures caused by outbound relay restrictions on shared hosting.
- Reconfigured secure mail delivery using provider-specific SMTP relay instead of generic mail host.
- Repaired domain-level email authentication (SPF + DMARC) to restore full mail deliverability.
- Debugged production-only issues masked during local development (SMTP, SSL, cache, linter constraints).
- Engineered around legacy WebCentral CSS limitations while preserving modern responsive behaviour.
- Stabilised contact-form infrastructure for production reliability under real-world security conditions.
Tools & stack
- Frontend: HTML5, CSS3, JavaScript (vanilla)
- Media processing: FFmpeg (hero video animation, mirroring, stacking, compression)
- Backend: PHP (contact form handler)
- Email delivery: PHPMailer, SMTP authentication (SSL/TLS)
- Hosting & deployment: WebCentral cPanel, File Manager, password-protected staging
- Mail security: SPF, DMARC, authenticated SMTP relay
- Workflow & QA: VS Code, browser dev tools, hard refresh & cache busting

This project delivered a fully responsive, production-ready scientific website with secure email handling, optimised media performance, and a WebCentral-compliant deployment pipeline, providing Data Biome with a reliable digital foundation for public communication and future growth.

Learning

CodeCademy Career Path - Data Science / Analytics

The syllabus of the CodeCademy Career Path course “Data Scientist - Analytics Specialist” is listed here (CodeCademy).
This long thorough course involved theory, practical, exams, and multiple projects. I was able to learn SQL, Python 3, and Tableau as well as revise my knowledge on statistical analyses and data visualisation.
I was able to apply my newly acquired Python skills to analysis the largest proteogenomics dataset to date (DOI: 10.3390/ijms25168614 ).
All the projects of the CodeCademy course were guided except for the last one which could be anything I wanted.
For the final portofolio, I chose to analyse the metadata of 120,000 scientific articles (dataset retrieved from Kaggle) to understand publication trends and topics of interest.
workflow icons

1. Hypotheses
- Publication rate over time
- Successful publishers and favored journals
- Prolific authors
- Number of citations and length of title over time
- Preferred publication types
- Languages used
- Most popular subjects

Link to GitHub repository containing the raw and clean datasets, as well as the Jupyter notebook with detailed Python code

2. Data handling and analysis
Tools used: Excel, Python, Tableau

Data assessment helped decide which variables to keep, what needed transformation, and the existence of missing values.
Data wrangling and cleaning involved coercing dates, renaming headers into meanigful names, replacing missing values, transforming article titles and creating new numerical variables for data analsyis.
Exploratory Data Analysis (EDA) included line charts, wordclouds, bubble plots, bar and pie charts, treemaps, scatterplots, histograms and box and wisker plots.

3. Data visualisation, interpretation and storytelling using Tableau and Powerpoint
Link to Tableau Public dashboard and data analysis
Report on my process, analysis and observations

4. Conclusions
- Yearly rate of publication gradually increases from the 60’s onward to reach 2,500 publications per year in the mid-90’s.
- 2007 featured many more publications than other years. Is it a dataset bias or did something happen to explain such spike?
- Most publications are journal article (94%,) followed by proceeding articles (3%), and book chapters (2%)
- The popularity and longevity of journals vary immensely
- Most successful publishers host a larger number of journals
- Publications are overwhelmingly written in English
- The number of citations as well as the title length are both on the rise
- The most popular subjects deal with social sciences, computer science, electrical and general engineering as well as chemistry
- Some subjects are long-lasting while others are short-lived or emerging
5. Future directions
- Create more sophisticated visualization (e.g. Sankey or chord diagrams) to better explore relationship across variables
- Refine the text-mining models using other LNP methods
- Fetch more metadata (co-authors, keywords, geographic data,...)

CodeCademy Career Path - Data Science / Machine Learning

The syllabus of the CodeCademy Career Path course “Data Scientist - Machine Learning Specialist” is listed here (CodeCademy).
This long thorough course involved theory, practical, exams, and multiple projects. I learned about Machine Learning algorithms via a foundation of math, stats, and programming (Python). The course covered supervised and unsupervised learning algorithms to solve classification and regression problems. Such algorithms were Ensemble methods, SVMs, Recommender Systems, Naive Bayes, Deep Learning and Neural Networks.
For the final portofolio, I was tasked to analyse a dataset from OK-Cupid Date-a-Scientist app and apply ML algorithms in any way I saw fit.
I chose to develop an unsupervised ML model that would identify individuals with similar features so that the app can suggest relevant matches.
The results will then be displayed via an interactive interface.
brainstorming icons

1. Study design
- Context
- Rationale
- Tools
2. Data handling, encoding and distribution using Python
Link to GitHub repository containing the cleaned and encoded dataset, as well as the Jupyter notebook with detailed Python code
- Data inspection and visualisation helped determine data types, cleaning steps, and the existence of missing values, erroneous values and outliers.
- Data wrangling and cleaning involved filtering and normalising numerical variables, encoding categorical ordinal and nominal features, and processing textual features using NLP methods for data analsyis.
- Plotting the data distributions as histograms, pie charts, treemaps, wordclouds, heatmaps and bar plots helps summarise the features and reveal the main trends.
- From those main trends can be inferred which types of individuals (archetypes) are interested in using the dating app.

3. Data analysis, machine learning, and data visualisation using Python
- Correlation matrix reveals no collinearity and associations between fetaures useful for NAs imputation.
- Conctenation and classification of textual features using Latent Dirichlet Allocation (LDA) with optimised parameters outlines 30 main themes.
- Features are weihgted according to their relevance for matching individuals.
- Row selection is based on a completeness score reflecting the frequency of NAs across all weighted features.
- Identification and elimination of outliers to reduce noise using Isolation Forest and optimised parameters.
- Grouping of individuals to decomplexify data using HDBSCAN clustering with optimised parameters and UMAP plot shows a structure of the data into 5 unven groups.
- Pair-matching of individuals within HDBSCAN clusters using Cosine similarity scores is achieved at a F1 score of 90%.
- An interactive interface using Gradio was developped to display up to 20 closest matches per matched individual.

4. Interpretation and storytelling using Tableau and Powerpoint
Link to Tableau Public dashboard and data analysis
Report on my process, analysis and observations

5. Conclusions
- EDA revealed the archetypical dating app user.
- Several ML models were implemented to categorise textual features (LDA), eliminate outliers (Isolation Forest), structure the data (HDBSCAN clustering) and pair the most similar individuals (cosine).
- Visualisation of the closest matches using heatmap, worcloud, and interactive tables help outline the general trends of selected individuals.
6. Future directions
- Different encoding of the features (e.g. agglomerate some variable labels to decrease granularity).
- No weighting of the variables (no bias) or different weighting strategy.
- No row selection, instead testing different NA imputation methods (mean, KNN,…).
- No outlier elimination, instead analyse them and associate them to features
- Use UMAP projections as predictive features in cosine model for improved granularity
- Retrieve GPS data of the town for geographical mapping to then including geographical proximity to the model.
- No combination of the 10 essay variables prior to NLP modeling.
- For NLP analysis, compare LDA method to BERT modeling which incorporates context as well.
- Test model with new users (fictional or retrieved from public repository).
- Further develop the Gradio interface to search for traits (e.g. “wants more kids”) rather than profile id.

CodeCademy Skill Path - Build a website

The syllabus of the CodeCademy Career Skill course “Build a website with HTML, CSS and GitHub pages” is listed here (CodeCademy).

Having completed this short course, I set out to create my professional website (the one you're browsing!).

I took this opportunity to create my brand. I opted for a simple and clear design with vivid contrasting colors. I display my full name in the header and my initials in the browser tab (favicon).
coding language icons
I used the Visual Studio (VS) Code IDE with the Live Server extension to create, edit, test and debug the Hypertext Markup Language (HTML), javascript (JS), Cascading Style Sheet (CSS) codes.
When statisfied with the result, I published my website using GitHub desktop.
coding language icons
The navigation bar allows for quick access to the various website sections. It stays on top and is transparent so it doesn't obscure what lays under it. It also toggles for small screens.

Sections are scrollable to minimise vertical layout. Hidden contents visible via clickable tabs allow for a succinct yet interactive display.
coding language icons
Clickable items (tabs, URLs) change color for better interactivity. Documents can easily be accessed and open in a new tab.
I've added lots of images to make it more enticing and to better showcase my skills. I used flexbox to optimise element positioning.
I've included images description via ARIA alt attribute and presentation role to make my website accessible to visually-impaired users.
coding language icons
The responsive design should maintain readability and visual appeal across all devices, and screen sizes.
I keep on updating my website so that viewers can keep track of my latest skills, projects, and publication.
I hope you'll find my website user-friendly and informative!

CodeCademy Skill Path - Build a chatbot in Python

The syllabus of the CodeCademy Career Skill course “Build a chatbot in Python” is listed here (CodeCademy).

The topics covered are:

Rule-Based Chatbots
Retrieval-Based Chatbots
Deep Learning and Generative Chatbots

The course includes Python, Natural Language Processing (NLP), Data Science, Machine Learning (ML), and Artificial Intelligence (AI).

Projects completed in this course are:

Various chatbots (GitHub)
Machine translation (GitHub)
Twitter (GitHub)

My capstone project is to develop a chatbot for my professional website.
The various steps include:

Develop a custom Python backend using FastAPI and deployed it on Render to handle chat requests securely and efficiently.
Integrate a JavaScript chat widget into this GitHub Pages site for seamless, real-time interaction with visitors.
Implement multi-stage intent recognition:
- Rule-based patterns for navigation and commands,
- Fuzzy string matching with RapidFuzz to interpret typos and informal phrasing, and
- FAQ retrieval from a structured faqs.json file for content-specific answers.
Add a lightweight TF-IDF semantic search engine that indexes the main sections of the website (Biography, Skills, Portfolio, Certificates) and provides concise, content-aware answers with direct “Read more” links.
Create an admin /reindex endpoint to hot-reload FAQ data and rebuild the TF-IDF site index without redeploying the app.
Design a privacy-first contact workflow that masks any personal information and invites genuine leads to connect via the embedded CV link.
Configure automatic deployment and updates through GitHub commits and Render auto-builds, ensuring continuous delivery of improvements.
Structure a clean, modular codebase ready to evolve toward retrieval-augmented generation (RAG) or LLM-powered dialogue in the future.

The chatbot you see here now combines rule-based logic, fuzzy matching, and semantic retrieval to provide context-aware responses.
Future enhancements will include conversational memory, dynamic topic expansion, and integration with generative AI models for richer, adaptive dialogue.
You'll find my chatbot widget at the bottom right corner of my website. Look for the icon below, click on it and try a query like "What is your okcupid project about?".

Biography

Research Scientist (1998-2023)

Data Scientist (2024-present)

Work with Me

Research Scientist (1998-2023)

Academic and Research Career

Quantitative and Statistical Expertise

Bioinformatics and Scientific Data Interpretation

Scientific Communication and Leadership

Data Scientist (2024-now)

1/ CodeCademy Career Path: Data Scientist - Analytics Specialist

2/ CodeCademy Career Path: Data Scientist - Machine Learning

3/ CodeCademy Skill Path: Build a Website with HTML, CSS and GitHub Pages

4/ CodeCademy Course: Learn Prompt Engineering

5/ CodeCademy Skill Path: Build a Chatbot with Python

6/ CodeCademy Course: Large Language Model (LLM)

7/ Automating Visual Storytelling with Python

Work with Me

Independent Consultant

How I Can Help

Typical Deliverables

Suitable Projects

Contact Form

Skills

Computer OS

Programming

Statistical Analyses, ML and NLP

Bioinformatics and Data Mining

Data Visualisation

Data interpretation and Storytelling

Editor and Reviewer

Computer OS

Programming languages, IDE, cloud platforms and AI

Statistical Analyses, ML and NLP

Bioinformatics and Data Mining

Data Visualisation

Data Representation, Interpretation, and Storytelling

Editor and Reviewer

Portfolio

Systems Biology & Post-Genomics

Big Data

Machine Learning and NLP

Patents

Visual Automation

Website design

Systems Biology and Post-Genomics Projects

Big Data Projects on Wheat

Machine Learning and NLP

FNRL Mode-of-Action in Arabidopsis Roots: an Integrated Multi-Omics, Machine Learning and Bioinformatics Study

Machine Learning Reveals Synergistic Effects of Lactobacillus helveticus in Camel Milk Fermentation Using Metabolomics

Machine Learning Modelling of Synergy and Antagonism of Phenolic Compounds Using Antioxidant Assays

Bread Protein Biomarker Discovery Assisted by Machine Learning

Career Trajectory Mapping From My Scientific Publications Using NLP, Machine Learning and Analytics

Aliens, Algorithms & Anomalies: Visualising the NUFORC UFO Sightings

Decoding Love with Data: Matchmaking in a Dating App

Mapping the Landscape of Scientific Publications

Patents

Visual Automation

Website design

Learning

CodeCademy Career Path - Data Science/Analytics

CodeCademy Career Path - Data Science/Machine Learning

CodeCademy Skill Path - Build a website

CodeCademy Skill Path - Build a chatbot in Python

CodeCademy Course - Learn Prompt Engineering

CodeCademy Career Path - Data Science / Analytics

CodeCademy Career Path - Data Science / Machine Learning

CodeCademy Skill Path - Build a website

CodeCademy Skill Path - Build a chatbot in Python

CodeCademy Certification - Learn Prompt Engineering

Certificates and Diploma

Research Scientist
(1998-2023)

Data Scientist
(2024-present)

Work with
Me