Peyman
Rasouli

An explainability layer built into AI-driven vaccine design pipelines at a biotech company developing personalised cancer vaccines. Clinical researchers and regulatory reviewers needed genuine insight into why a model ranked certain candidates — not just the prediction, but the reasoning behind it. Delivered interactive dashboards and prototype-based transparent models that made AI behaviour visible and auditable to domain experts.

Fine-tuned protein language models for classifying biological sequences in the context of personalised cancer vaccine development — where off-the-shelf models fall short of clinical and scientific standards. The work involved base model selection, rigorous evaluation framework design, and systematic benchmarking against state-of-the-art to validate every training decision. The result was a production-grade classifier that measurably outperformed baselines and directly informed the vaccine development roadmap.

An autonomous AI pipeline for identifying vaccine target candidates — combining semantic search over biological literature with sequence alignment and explicit human escalation triggers. Designed around uncertainty: confidence thresholds determined when the agent could proceed and when a domain expert needed to intervene. The result was a faster, more systematic approach to vaccine target discovery that kept human judgment in the loop where it mattered most.

A retrieval-augmented generation system producing AI outputs calibrated and interpretable enough for regulatory review in a clinical setting. The challenge was not retrieval accuracy alone — regulatory audiences need to trace the reasoning behind every conclusion drawn from complex biomedical literature. Delivered a system combining structured retrieval with rigorous benchmarking, meeting the interpretability bar that regulated healthcare environments require.

End-to-end ML models predicting missing petrophysical curves and lithology from well log data, deployed into the daily workflows of geologists and petrophysicists at an oil & gas client. The work covered data engineering, model development on cloud platforms, and integration into the client's industrial data platform. The measure of success was adoption — geoscience teams relying on the predictions as a standard part of their operational workflow.

A streaming ML system predicting subsurface properties on the fly during active drilling operations — processing live sensor data and returning predictions fast enough to inform in-the-moment decisions in the field. Built for a safety-critical environment where latency and reliability are non-negotiable. Delivered AI-driven intelligence embedded directly into drilling operations, putting data-driven decisions in the hands of geoscientists and drillers in real time.

A research project combining domain knowledge graphs and ontologies with ML explainability methods — making model explanations grounded in the structured knowledge that domain experts actually use. Standard XAI methods produce explanations that are statistically valid but often semantically meaningless to practitioners; this work addressed that gap directly. The result was an explainability approach that spoke the language of the domain, not just the model.

A methodology and toolkit for validating explainability methods in real industrial deployments — bridging the gap between research algorithms that perform in controlled settings and the practical demands of healthcare and energy environments. Designed evaluation frameworks, worked directly with domain practitioners to assess explanation usefulness, and iterated on what actually helped people make better decisions. Released as open-source tools adopted by both research and industry communities.