Agentic workflow architecture
Human-gated workflows for AI-assisted data science: scoped context, controlled execution, review gates, reproducible outputs, and clear handoff.
The real problem
Health data work needs more than automation.
Many teams want the speed of AI and coding agents, but their work depends on sensitive data, domain context, statistical assumptions, and decisions that cannot be delegated blindly. The question is not simply whether to use AI. The question is how to design the workflow boundary: what agents can see, what they can do, where review happens, and who remains accountable.
A different approach
Agents can build. Experts still decide.
I design workflows where agents assist with structure, coding, retrieval, reporting, triage, and documentation, while statisticians, analysts, scientists, clinicians, or domain experts remain responsible for the question, the assumptions, the evidence standard, and the final decision.
Execution from agents. Judgment from humans.
The goal is not to remove people from data science. It is to give them safer systems, clearer gates, reproducible evidence, better defaults, and more time to think.
Services
Four ways to work with me.
Biostatistics and health-data review
Study design, analysis plans, modelling strategy, assumptions, uncertainty, and interpretation for health-related data and decision support.
Reproducible R/data systems
Reusable data cleaning, validation, reporting, packages, dashboards, and lightweight tools your team can run, inspect, and maintain.
Data and system safety
Practical boundaries for agentic work: data minimization, local/remote separation, permission levels, documentation, machine review, and human approval.
What I build
Transparent systems your team can keep.
Everything I build is designed to be read, understood, validated, rerun, and maintained — not hidden behind a black box.
- Agent-assisted analytic workflow designs
- Human review gates and release checkpoints
- Reusable data cleaning and validation pipelines
- R packages and internal utilities
- Quarto reports and automated reporting workflows
- Shiny and lightweight web tools
- Spreadsheet-to-pipeline transitions
- Documentation, SOPs, and maintainer handoffs
My philosophy
Biostatistics grounds the architecture.
In health-related data work, a workflow is useful only if people can understand what it is doing, why it is doing it, where it might fail, and who remains accountable for the decision. AI can help with execution, but it should not hide assumptions, uncertainty, or responsibility.
Humans frame the questionAgents can assist the work, but people define the decision, context, and success criteria.
Assumptions stay visibleModels and automation are useful only when limits, uncertainty, and failure modes are clear.
Sensitive context is minimizedGood systems expose only what is needed and keep sensitive material behind the proper boundary.
About
Lennon Li
I am an agentic workflow architect with expertise in biostatistics, public health, and health-related data. My background combines statistics, computer science, applied public health, R/data systems, and machine-assisted decision making.
I have spent more than 15 years building statistical analyses, surveillance models, interactive tools, reproducible workflows, and decision-support systems for teams working with complex and sensitive health data.
My focus now is trustworthy agentic data science: designing workflows where machines help with planning, code, retrieval, triage, reporting, and documentation, while humans remain responsible for assumptions, meaning, interpretation, and final decisions.
Process
How I work
- Understand the decision — what question, decision, or workflow will this system support?
- Set the data boundary — what agents may see, what must stay local, and what can be safely summarized or simulated?
- Design the gates — where machine output needs review, approval, escalation, or rejection.
- Build the workflow — code, model, report, package, app, or tool is developed as a reusable system, not a one-off artifact.
- Validate and hand over — assumptions, outputs, failure modes, documentation, and maintainer steps are made explicit.
The shift
From “use AI” to “design the workflow.”
Running an agent may solve one task. Designing a human-gated workflow makes the next task safer, faster, and easier to review.
Contact
Need safer agentic data-science work?
If your team works with health-related data, repeated reports, exported files, manual analysis steps, fragile pipelines, or AI-assisted coding workflows, I can help turn that process into something secure, reproducible, sustainable, and human-gated.
Please do not send sensitive data by email. We can first discuss the workflow and decide on an appropriate secure process.