01 / Quantify
Quantify evidence
Statistical models turn observations into explicit estimates, forecasts, comparisons, and uncertainty. Assumptions and validation remain visible rather than disappearing behind a single score.
BIOSTATS + AI · DECISION SCIENCE · HUMAN JUDGMENT
Augment Intelligence. Retain Accountability.
Ye (Lennon) Li PhD · P.Stat.
Biostats + AI combines statistical rigor and machine intelligence to strengthen evidence-based human decisions. AI supports reasoning and evaluates alternatives; people retain final decision authority and accountability for consequences.
evidence / quantitative + semantic
uncertainty: carried forward
model authority: bounded
decision: human-owned
One reality. Many perspectives.
INTERACTIVE / 30 SECONDS
A prediction is not a decision.
Change the probability, sensitivity range, and costs of error. See when the preferred action changes across assumptions; a threshold crossing does not establish that review is optimal.
ADJUST THE ASSUMPTIONS
ILLUSTRATIVE DECISION EVIDENCE
Plausible range30–54%
Action threshold40.0%
0%25%50%75%100%
Illustrative rangeEstimateThreshold
What this suggests
Human review needed
The illustrative uncertainty range crosses the threshold; human review is needed.
Illustrative guidance: Human review neededHUMAN DECISION / REVIEWABLE CHOICE
The system's guidance is not the final decision.
Choose an action and optionally record why. This educational demo shows your choice locally in the page; it does not store or transmit your input.
No human decision recorded yet.
Investigation is not an optimized third action here: its value depends on investigation costs, information reliability, and the subsequent decision it could change. A deployed system would also track authorization, outcomes, and retrospective evaluation.
Educational example, not a recommendation for any real-world decision. Values and the ± range are invented, not statistically calibrated predictions or confidence intervals. You set the relative missed-event cost; a fixed false-alarm cost of 2 yields the expected-loss threshold. Real decisions require validated evidence, context, and human judgment.
01 — From evidence to decision
Most real-world problems become decisions.
Should we investigate, release, delegate, disclose, continue, or stop? Classification and prediction inform these choices, but action also depends on alternatives, constraints, uncertainty, costs, and consequences.
02 / Interpret
Interpret and reason
Not every useful input is numeric. Semantic models interpret context; interactive LLMs help frame questions, surface competing explanations, challenge assumptions, and translate evidence without acquiring decision authority.
03 / Decide
Evaluate the decision
Compare actions, asymmetric errors, opportunity costs, and the value of more information. Evaluate whether AI improves final decisions over simpler baselines; people retain authority and accountability.
Operating principle
Uncertainty should travel with the answer.
The system should preserve enough evidence for a person to understand what supports a recommendation, what remains unknown, and when not to act.
Data are projections of reality.Models are projections shaped by data and assumptions.Reality is larger than both.
02 — Decision-system method
Frame, deliberate, decide as a human—and evaluate consequences.
The method travels across privacy, scientific review, investigation, model delegation, resource allocation, and stopping decisions. An LLM supports iterative reasoning and explanation; calculations, constraints, and validation remain inspectable.
01
Decision
What choice, escalation, forecast, review, or action does the system need to support? What would count as useful evidence?
defined02
Evidence
What quantitative and semantic evidence is available? What is missing, biased, unsafe to expose, or outside the observed data?
bounded03
Models + uncertainty
Use the right statistical, computational, or semantic models; validate them; compare alternatives; and carry uncertainty, disagreement, and abstention forward.
evaluated04
Deliberate and evaluate
Use interactive AI to question assumptions and explain trade-offs; compare costs and consequences, keep the human decision explicit, then test decision quality, calibration, and review burden against baselines.
human-ownedDECISION LOOP / HUMAN IN THE SYSTEM
The human is part of the system—not an approval button outside it.
Statistics and AI provide evidence, reasoning, alternatives, and consequence analysis. People define goals and value trade-offs, challenge assumptions, decide whether to act, investigate, defer, or stop, and remain accountable for their decisions and consequences. The system should preserve the rationale, human authorization, feedback, and opportunities to reassess.
Accountability is shared, not transferred: developers, deploying organizations, and governance bodies also retain responsibility for design, validation, oversight, and conditions of use.
Better automation is useful. Better judgment and clear accountability are the objectives.
RESEARCH QUESTION / NOT YET VALIDATED
Can biostatistics, AI, and human judgment together improve decisions?
This is a hypothesis to test, not a result already established. Existing tools and interactive examples demonstrate methods and functionality; they do not yet establish improved real-world decision outcomes.
BUILT
Methods and prototypes
Statistical models, software tools, rules and semantic processing, and interactive hypothetical decisions. Demonstrated functionality is not demonstrated effectiveness.
TO BE EVALUATED
Compare against simple baselines
Compare rules, statistical models, and AI-assisted approaches on calibration, decision loss, cost, review burden, and human-system performance.
NOT YET KNOWN
Real-world decision value
We do not claim AI has improved actual decision outcomes. Positive, negative, and inconclusive findings should all be reported transparently.
03 — Existing work, one theme
Decision science in practice.
I explore the same questions across public-health investigation, scientific software, privacy and AI access, and efficient model delegation. These are domain-specific research and engineering examples—not claims of a universal validated decision engine.
Flagship / PAGe
Forecast what is still unfolding—and keep the uncertainty visible.
PAGe is a phase-aligned infectious-disease forecasting system. It treats forecasting as a decision-support problem: detect the emerging phase, estimate what may happen next, preserve probabilistic uncertainty, and provide evidence a person can interpret rather than a single opaque prediction.
Quantitativestatistical forecasting
Probabilisticuncertainty retained
Validatedout-of-sample evaluation
Humandecision support, not autopilot
PAGe v3 is shadow-only, not an operational clinical forecasting service.
PROJECT PROOF
Four projects. One evidence-to-decision approach.
Scientific forecasting, decision gating, privacy-aware AI, and reproducible engineering. Inspect working evidence and the methods behind each system.
01 / PROJECT WORK
PAGe · Scientific forecasting
Forecasting one and two weeks ahead with expert-reviewed epidemic ignition, staged model training, frozen artifacts, and holdout checks. The current v3 is shadow-only.
CAPABILITY DEMONSTRATED
Scientific modeling · gated evaluation · reproducibility
02 / DECISION GATING · RESEARCH PROTOTYPE
Cranary · Evidence before release
An evidence-backed R package screener that separates deterministic checks, bounded semantic review, and human release decisions. A finding can require action, review, or remain unknown; AI cannot override an unresolved rule or certify CRAN acceptance.
BENCHMARKING THE GUARDRAILS
Historical and current CRAN package versions, documented issue→fix pairs, and held-out controls test whether risk findings are useful and change when defects are resolved. Longitudinal scores and wider package benchmarks are being developed—not yet validated as general rankings.
Active development · source currently privateEvaluation method ↗
03 / PROJECT WORK
DataGangeR · Safer AI development
An R package that makes local synthetic stand-ins for data-sensitive workflows. Users review sensitive fields, fidelity, and disclosure warnings before export. No anonymity guarantee.
CAPABILITY DEMONSTRATED
Statistical engineering · privacy review · agent handoff
04 / PROJECT WORK
ONgeoR · Geography you can audit
Retrieves authoritative Ontario and census boundaries, links locations to public-health regions, records sources, and offers a Shiny interface that exports reproducible R code.
CAPABILITY DEMONSTRATED
Geospatial engineering · provenance · usable tools
Project descriptions reflect their published documentation. Operational fitness requires independent evaluation.
Decision systems
evidence → actionPAGe supports infectious-disease forecasting. Which explores explicit decision and evaluation layers for calibration, abstention, model comparison, and semantic evidence. aiR supports statistical analysis while keeping assumptions and interpretation with the analyst. The SSC Review Agent applies retrieved evidence and structured rubrics while leaving final decisions to human reviewers.
Trust + control infrastructure
bounded AIDataGangeR reduces unnecessary exposure of source data through synthetic stand-ins. HMA binds evidence and approvals to workflow stages. AgentPorter constrains what agents can access and execute. clientAgent turns requests into reviewable candidate artifacts without giving agents release authority.
Reproducible evidence systems
provenance + validationOPCC and ONgeoR make public-health linkage and geography reproducible and source-documented. Cranary explores evidence-backed package screening, bounded AI review, and human-gated release decisions, using CRAN package-version benchmarks to test its findings. These systems make the evidence trail inspectable and reusable instead of leaving it implicit in one-off analysis.
AI expands what can be represented—not what should be trusted automatically.
Statistics contributes explicit assumptions, calibration, validation, and uncertainty. Semantic AI contributes a way to work with evidence that is difficult to encode as conventional variables. Workflow controls preserve provenance, boundaries, and review.
The common design question is therefore not Which model is smartest? It is What evidence supports this decision, what remains uncertain, when should the system abstain, and who has authority to act?
evidence?calibration?uncertainty?abstain?provenance?authority?
04 — WORKING SOFTWARE
Try the tools—not just the descriptions.
Public demos of statistical workflows and decision-facing interfaces. Use synthetic or public-safe data only.
Data boundary: Never upload personal, patient, confidential, or client-owned information to public demos. Read the privacy notice. ONgeoR: source · documentation.
HOW I THINK
What the models cannot see.
Biostats + AI joins quantitative discipline with new ways to interpret and discuss evidence. Daoist and Buddhist thought remind me that every model is partial, conditions change, and confident descriptions are not reality.
A useful model is not the whole truth. A convincing AI answer is not a decision.
My work keeps the trail from observation to interpretation visible: evidence, assumptions, uncertainty, privacy boundaries, and human responsibility.
What remains unseen
The reality that can be modeled is not the whole reality.
The point is not to eliminate uncertainty. It is to see more clearly, know the limits of what we see, test alternatives, recognize bias, and remain open to what our models have left out.
Observe carefully. Model humbly. Decide consciously.