AIthena: The Data Foundation
Clinical trial data treated as a product, so the agents have something clean to work on
The data layer under Navari and Agentis. AIthena is an operating model that makes clinical and translational trial data AI-ready at source, with an accountable owner and release gates per study. Its first proof point was a ctDNA atlas across 24 trials and 4 vendors, concept to MVP in two months. I sponsored and shaped it; a program lead ran it day to day.
The problem
The Data Supply Chain
Trial data arrived raw, study-specific and vendor-specific, and got curated by hand months after the fact. Dozens of hand-offs sat between a dataset and a model that could use it. Everyone was working on the end of the pipeline, AI consumption, while the beginning, how data gets generated and contracted, stayed as it was.
- Weeks to months before new trial data was usable for analysis or AI
- Curation done retrospectively, by hand, per study, often by a contractor with no program authority
- Accountability for data quality split across study teams, data management and vendors
- Agents and models starved of clean inputs no matter how good they were
What I built
Shift Left on Data
Treat clinical and translational data as a regulated product with explicit release criteria. Fix the beginning of the pipeline and the middle gets cheap. AIthena is not a new tool to install; it is an operating model with named roles, gates and contracts.
- A data pod per study, with a named owner accountable for release and quality
- Release gates from study design through close: standards set in the biomarker plan and the vendor contract, not patched on later
- Vendor data delivered to specification, with acceptance criteria that have teeth
- FAIR at source: cataloged, API-accessible, on common data models, reusable with lineage
How it fits together
Architecture
Accountability is embedded in each study, standards are enforced at contracting, and a central catalog and API layer make the resulting data products findable and consumable by people and agents alike.
Timeline
How it went
- Q1 2025Setup: data pods staffed, standards defined, vendors onboarded
- Q2 2025Pilots on new trials, with external vendor dependencies on purpose
- Q3 2025Operational, a quarter ahead of plan
- 2025ctDNA atlas: 24 trials, 4 vendors, first internal data agreement of its kind, concept to MVP in 2 months
- Apr 2026Planning the scale-up from pilot to enterprise
- Jun 2026Handed off; executive sponsorship transferred
What it taught me
Lessons
- The model was never the bottleneck. The data supply chain was.
- Fix the beginning of the pipeline and the middle gets cheap. Retrospective curation never catches up.
- Accountability per study beats a central data team trying to clean everything after the fact.
- A contract with acceptance criteria does more for data quality than any pipeline.
- Sponsor it, shape it, then get out of the way of the person running it.
Stack
Connected work