Tag.bio data products are the fundamental building block of the data mesh architecture and implement a domain-driven design to ensure that each data product is usable by anyone - human or AI.
Data products represent a harmonized, composable application layer on top of disparate, multimodal data sources. Along with employing a universal “Smart” API, they also present a simple, clean, standardized data model for apps, data scientists, and AI agents who perform queries and extract data frames. Data products have the following features:
Each data product presents its data model to AI, algorithms, BI tools, and data scientists as a single, sliceable data frame with common vocabularies and ontology concepts.
Disparate data products can map into an archetypal data model, which enables instant transfer of algorithms and data consumption methodologies.
All data snapshots, mapping, statistics, ML models, visualizations and reports are versioned for analysis provenance and perfect reproducibility.
Each data product utilizes its own git repository (or archetype repository) and is deployed as a Docker container into a robust, fully-managed Kubernetes compute infrastructure.
Data products can be designed, managed, and tested independently, minimizing overhead and maintenance costs as data and usage evolves.
Tag.bio's Data Mesh presents a network of interconnected data products, adhering to FAIR principles via smart APIs. This framework promotes data discoverability and interoperability, ensuring that both data and analyses are FAIR-compliant, expediting the process of connecting real-world data sources.
Turnkey archetypes for CDISC, OMOP, EHR, RWE, DNA-Seq, and RNA-Seq (including single-cell), plus low-code JSON/YAML modeling for anything novel, mean a new real-world data source becomes a governed data product the same day — not a multi-quarter integration project.
Each team owns and runs its own data products, while a shared set of smart APIs and standards keeps the whole mesh FAIR-compliant and interoperable — so domains move independently without a central team gatekeeping every connection.
Every analysis, cohort, and report run against the mesh is automatically saved and cataloged, so institutional knowledge accumulates for the next person to build on — instead of leaving when the person who ran it does.
.png)
Access is enforced at the data layer using each user's own credentials — not written into a system prompt — so an assistant only ever sees the data products, and the slices within them, that user is entitled to see.
A protocol runs over every row in the data product — tens of thousands of samples, tens of thousands of variables — and only the computed result comes back to the model. AI cost scales with the size of the answer, not the size of your data.
Every data product speaks the same API, so a question that spans a clinical trial, a reference cohort, and a handful of published studies runs as a single query pattern — not a new integration project for each source.
The data products, the protocols, and the MCP connector all run inside your own secure cloud account, on top of your existing storage — your data never has to leave your environment to be analyzed.
Model-agnostic, by design
Tag.bio doesn't lock you into one model. Point any foundation model — OpenAI, Claude, Gemini, or an open model like Qwen or DeepSeek — at a data product, and it can build, query, and reason over it the same way. Biology-specialized models get the same access: protein-structure models like AlphaFold3 and ESMFold, or biomedical language models like BioBERT and ClinicalBERT, can invoke the same protocols and ask the same ad hoc questions. The data product layer is the constant; the model is your choice.
How AI builds and uses data products
Feed AI your raw, disparate sources and it assembles the data product: mapping fields, applying a standard data model, versioning as the underlying data changes. What used to be a data engineering project becomes a starting point.
Once a data product exists, AI can invoke a protocol (a proven, reusable analysis) or run an ad hoc query directly against it. Either path stays inside the governed layer — there’s no separate export, no shadow copy of the data floating around in a notebook.
Protocols
A protocol is a defined, reusable analysis that runs inside the data product layer instead of inside your AI’s compute budget. Encode a validated analysis once, and every future request for it — whether it comes from a person or an AI agent — runs the same way, on the data product itself. That means faster answers, lower AI compute costs, and one less place for inconsistency to creep in.
Agentic AI
As AI moves from answering one question to running a chain of them, that chain needs to run against something governed. Tag.bio's data products are built to be that ground: autonomous agents connect through MCP, invoke protocols or run ad hoc queries, and hand results to other agents — including agent-to-agent communication across a federated network of agents working the same data product layer. Retrieval stays grounded in the data product itself, not a loose index of exported files, so an agent chaining five steps together is still logged, auditable, and tied to the data it actually touched at every step.
MCP
Tag.bio can front all of your data products with a single MCP server, so any AI system that speaks MCP — including Claude — can discover and query them directly. Every connection, every query, is scoped and logged.
Want to see it before you build anything? We host a live MCP server at demo.tag.bio that you can connect to today with a free Claude connector (limited usage) — a real data product, not a sandbox with fake data.
Reproducibility and auditablity
Anytime AI interfaces with a data product — a protocol run or an ad hoc query — it’s logged: the user, the request, the result. That history is auditable and reproducible by design, which matters as much for a regulatory review as it does for a colleague trying to pick up where you left off.
Because the data product sits on top of your existing storage infrastructure rather than duplicating it, it becomes a single ground truth — the version of the data that any analysis, dashboard, or AI agent can be pointed at with confidence.
