Implemented Solutions
QSAR Model Development
Building reproducible cheminformatics pipelines for QSAR modelling and model validation.
In collaboration with
Technical Details
Implementation Specs
- Descriptor Calculation: use of AlvaScience for molecule dataset cleaning and standardization, molecular descriptors calculation, and fingerprint calculation.
- Feature Selection: filtering methods are combined with embedded method to reduce to the dataset to the most important features for the endpoint.
- Models: different models tested and compared for classification task. Spanning from LDA, QDA to more complex tree-ensemble algorithms such as Random Forest.
- Validation: cross-validation and external test set validation.
- Deployment: models used in software for QSAR.
Manufacturing Optimization
Monitoring failure rates and identifying critical trends in production data via chemometric analysis.
Technical Details
Implementation Specs
- Data recollection and cleaning: standardization of the data and handling of missing values, incorrect values etc.
- EDA pipeline: data analysis using basic statistic and multivariate analysis such as PCA for batch evolution modelling.
- Root Cause Analysis: identification of critical process parameters affecting out-of-specification values of a variable in API synthesis.
- Results: data-driven decision to address next experiments and steps to solve the problem.
QSRR for Chromatography
Building ML pipelines for retention time prediction using standardized chemical structures.
PhD research at
Technical Details
Implementation Specs
- Dataset Curation: building and curation of datasets of molecules, chromatography and mass spectrometry data, and molecular descriptors.
- Feature Selection: development of a feature selection pipeline to ensure reliable and significative results, combining filtering and embedded techniques.
- Algorithm: different algorithms including ensemble learning (XGBoost) optimization for non-linear chromatographic relationships.
- Results: better understanding of the retention mechanism and reduction of false candidate for MS unknown analyses.
AI for Wet-Lab Scientists Training
Lecture taking doctoral researchers from clean chemical data to deployed AI tools.
Delivered for
Technical Details
Course Outline
- Delivered to the SEACHEM and INCLUE MSCA doctoral networks: a 4-hour course across 3 modules, from clean data to deployed AI tools.
- Module 1 - Data Cleaning: building a clean data foundation through collection, standardization, and unit and naming harmonization.
- Module 2 - Data Analysis: EDA and multivariate analysis (PCA), the supervised ML workflow, and models for yield, failure rate, and root-cause analysis.
- Module 3 - Cheminformatics: turning molecules into descriptors and fingerprints for QSAR/QSPR/QSRR property and retention-time prediction, with applicability domain.
- Outcome: researchers able to build reproducible ML pipelines from clean data.
Secure Document Intelligence
An AI platform for querying confidential document archives such as regulatory dossiers, contracts, technical reports, with answers that stay verifiable, traceable, and fully inside your network.
Technical Details
Implementation Specs
- Data sovereignty: documents are indexed and queried by a local LLM, with no content ever leaving your infrastructure.
- Encryption: disk-level encryption (LUKS / AES-256), mutual-TLS between services, and vault-managed key rotation.
- Access control and traceability: role-based permissions (RBAC) with SSO and MFA, enforced per document collection, plus an immutable, audit trail.
- Input and stack integrity: Secure Boot / TPM verification, cryptographically signed updates, and document sanitization against prompt injection.
- Compliance by design: aligned with GDPR (verifiable right-to-erasure), NIS2, and ISO/IEC 27001.
- Scalable and modular: from a single GPU workstation to high-performance servers, with every component upgradable independently.