Every engagement follows five visible steps: know the data → sanity-check → build in an open stack → verify → report and hand over.
Work may begin from a published dataset, where the paper provides an answer key, or from your own data, where the checks are agreed from the design and data itself.
One accountable lead, with tools used where they help
A principal investigator, lab lead, or research-group lead can start with one scoped workflow; institution-level involvement is added where approval or wider rollout requires it.
Patrick scopes the work, makes the methodological judgement calls, reviews the evidence, and delivers the final output. Software and AI-assisted tools may help organise material, write or test code, compare results, and prepare documentation. They do not take academic responsibility or approve their own output.
A separate checking step is built into each engagement. What that check involves depends on the work: reproducing published statistics, testing code against known values, reviewing claims against sources, or checking a workflow against agreed acceptance criteria.
If you want the working environment for your own lab or project, see Own an AI Research Team.
The five steps
-
1
Know the data
What happens: For a replication, Patrick locates the public deposit and checks that it corresponds to the paper. For new work, he inventories the files, variables, codebook, provenance, and known exclusions.
Why it matters: The analysis needs a fixed starting point. For new data, the inventory also establishes what is present, what is missing, and what must be clarified before modelling begins.
-
2
Load and sanity-check
What happens: Row and column counts, variable types, value ranges, duplicates, missingness, and obvious inconsistencies are examined before the main analysis.
Why it matters: Many apparent analytical problems begin as data-definition or version problems. Showing these checks first makes the starting conditions visible.
-
3
Build or rebuild the analysis
What happens: For a replication, the core analysis is reconstructed from the paper and available materials. For new work, the analysis is built from the agreed research design in a scripted, documented stack.
Why it matters: The objective is a result that can be inspected and rerun, not an unexplained number copied from another output.
-
4
Verify
What happens: Replication results are compared with the published statistics item by item. New analyses receive checks appropriate to the design, such as diagnostics, sensitivity analyses, known-value tests, or review against agreed acceptance criteria.
Why it matters: Matches are recorded precisely. Differences, assumptions, and unresolved limitations are reported rather than smoothed over.
-
5
Report and hand over
What happens: You receive the agreed report, plain-language summary, scripts, and documentation, together with the relevant input and output files.
Why it matters: Your team can inspect the reasoning and rerun the delivered work after the engagement ends.
Before any data is transferred
The first scope includes a short privacy and data-handling pass. We identify:
- whether the proposed use is covered by the relevant consent, ethics approval, and institutional rules;
- which fields are necessary and whether direct identifiers can be removed;
- where files and working copies may be stored;
- who needs access;
- whether any external software or model provider would receive data;
- how long working copies must be retained; and
- when and how they will be returned or deleted.
No restricted dataset should be sent until that arrangement has been agreed. If the required safeguards cannot be supported, the work is redesigned for your lab's environment, research group, department, or institution, or not accepted.
What you keep — and where responsibility sits
You retain ownership of your data and research decisions. For commissioned work, you receive the agreed reports, scripts, documentation, and project-specific outputs, with reuse rights stated in the scope of work. Patrick retains ownership of pre-existing methods, templates, and general-purpose components unless the agreement says otherwise.
Patrick is responsible for the work and checks agreed in scope. You remain responsible for scientific conclusions, authorship decisions, ethics compliance, and any submission or institutional approval. No analysis can remove limitations in the design, measurement, sample, or available data; those limitations will be stated in the deliverable.
Pricing
Fixed-fee and scoped before work begins. Prices are stated in USD; HKD figures are approximate planning anchors.
| Tier | Price |
|---|---|
| Trial analytics deliverable | from US$1,000 (≈ HK$8,000) — scope and acceptance criteria agreed before work begins; fee waived if the completed deliverable does not meet them |
| AI research team setup | from US$2,000 (≈ HK$16,000) one time — separate fixed-fee engagement; not covered by the trial analytics assurance |
| Small deliverable | from US$2,600 (≈ HK$20,000) |
| Grant-scale project | from US$5,100 (≈ HK$40,000), normally milestone-based |
| Research retainer | from US$1,300 per month (≈ HK$10,000 per month) |
| Admin retainer | from US$640 per month (≈ HK$5,000 per month) |
The setup price does not include the server or third-party AI/model usage. Those recurring costs are paid by the client. Additional workflows, correction rounds beyond the included pilot, and ongoing support are quoted separately.
Whether an invoice is eligible under a grant or institutional budget depends on that funder's and university's rules. The scope and invoice can be written to document the service clearly, but approval rests with the relevant institution.
Choose the smallest useful starting point
For analysis or replication work, start with one fixed-scope trial analytics deliverable. Scope and acceptance criteria are agreed before work begins; if the completed deliverable does not meet them, the fee is waived.
For a lab-owned working environment, see Own an AI Research Team. The one-time setup includes one workflow pilot with agreed inputs, acceptance criteria, and one correction round.
For a larger or recurring requirement, request a scoped quotation. See the FAQ →