AI architecture in biotech becomes more useful when we stop treating “the LLM” as the system. A production system is a chain of boundaries: data access, retrieval, transformation, model reasoning, tool use, review, logging, and feedback. Each boundary is a place where quality can improve—or where trust can collapse.
1. Start with a decision, not a chatbot
The unit of design should be a workflow decision. What is the user trying to decide? What evidence do they normally inspect? What tools do they touch? What does “wrong” mean in this context? A system that summarizes documents and a system that influences an experimental decision require very different controls.
2. Build an evidence layer
Retrieval should provide more than relevant text. The system should preserve document identity, version, source, timestamp, and the specific evidence used to produce an answer. Provenance is not an audit feature added later. In research environments, it is part of the user experience.
3. Separate reasoning from tools
Models are good at choosing and sequencing actions, but deterministic work should remain deterministic. Database queries, calculations, schema validation, file transforms, and policy checks should live behind explicit tools with typed inputs and observable outputs.
The model should decide what to do only where probabilistic judgment creates value. Everything else should be made inspectable.
4. Evaluate the workflow
Generic benchmark scores are rarely enough. Build evaluations around the actual job: retrieval coverage, evidence correctness, tool-call validity, answer completeness, abstention behavior, and reviewer agreement. For high-impact workflows, test the failure modes you fear rather than only the happy path.
5. Keep a human control surface
Human review should not mean “read everything again.” Good control surfaces show the model’s evidence, the actions it took, the fields it changed, and the uncertainty that matters. The goal is selective attention.
6. Instrument adoption like a production system
Measure where users abandon the workflow, override the system, repeat queries, open evidence, or revert to manual processes. Adoption problems often reveal architectural defects: missing context, opaque outputs, too much latency, weak recovery paths, or a mismatch between the system and the real workflow.
The stack
- Workflow/API layer for identity, permissions, and orchestration.
- Evidence layer for indexed scientific and operational sources.
- Model gateway for routing, prompting, and policy.
- Typed tools for deterministic system actions.
- Evaluation harness for workflow-specific quality.
- Human review surface for evidence and interventions.
- Observability for latency, cost, failure, quality, and adoption.
That architecture is less glamorous than “put an agent on it.” It is also much closer to the kind of system I would trust to become part of serious scientific work.