Most AI in biopharma research stalls between capability and use. I close that gap by connecting scientific priorities, data, technology, products, and the people accountable for decisions, so AI becomes something researchers trust and reach for.
Four things I have come to believe about making AI matter inside a research organization.
Adoption is the hard part, not the model.
The capability usually already exists. What is missing is the path from a scientific question to a system a researcher trusts enough to act on — and that path runs through data, product, and organization, not through the architecture.
Measure decisions, not usage.
Dashboards count logins. The question worth answering is whether a decision was better, or arrived sooner, than it would have without the system. Everything downstream — investment, platform, partnership — gets easier once that is the standard.
Trust is earned on evidence scientists already accept.
Biological grounding and real clinical-trial decisions make a far better benchmark than a leaderboard score. A model that generalizes on paper and fails on the assets in front of you was never evaluated honestly.
The org chart is part of the architecture.
Where a capability sits, who owns the data, and who is accountable for the decision it informs shape outcomes more than model choice does. Getting that arrangement wrong is the most common reason good work stalls.
Stood up a dedicated AI team inside R&D and the operating model around it — how work gets selected, how scientists are brought in early, how delivery is judged. The output was never a set of projects; it was a capability the organization keeps using.
Defined the enterprise product strategy that moved R&D AI evaluation off usage dashboards and onto research impact, decision quality, and value. The decision framework it produced now guides AI investment, platform, and vendor priorities.
Built a translational strategy that used shared mechanism to work around sparse rare-disease data, integrating biological, clinical, and real-world evidence. Delivered as products scientists use, and applied across seven assets to prioritize repurposing and indication expansion.
Grounded model and vendor assessment in biological evidence and real clinical-trial decisions rather than leaderboard scores — five external technologies evaluated, two academic partnerships established, including Orphanet and the European Reference Networks.
Published in
The research record is where the standards come from — work on generalizability, negative sampling, disease representation, and how biological machine learning is measured. It is the same bar I hold products and vendors to.
Ravandi, C. B., et al. (2026). Clinical trial and ontology-derived positive and negative benchmark datasets for drug repurposing across rare diseases. Scientific Data (in review).
Ravandi, C. B., et al. (2026). A mechanistic disease representation framework for therapeutic knowledge transfer across rare and common conditions. bioRxiv, 2024.11.19.624381.
Ravandi, C. B., et al. (2025). Topology-driven negative sampling enhances generalizability in protein–protein interaction prediction. Bioinformatics, btaf148.
Ravandi, C. B., et al. (2025). Prevalence of processed foods in major US grocery stores. Nature Food.
Menichetti, G., Ravandi, C. B., et al. (2023). Machine learning prediction of the degree of food processing. Nature Communications, 14, 2312.
Full list on Google Scholar.
Advisory, board conversations, and talks. Boston-based, working globally.