The proof: your own published numbers, reproduced
Before any engagement, I re-run a researcher's own published analysis from their public data (OSF / PLOS / GitHub) and hand over a side-by-side comparison — your paper's statistics versus an independent re-run. No claims, no promises: you see your own results confirmed, or a discrepancy found — which is itself publishable as a reproducibility report.
| Method family | Worked example | Result |
|---|---|---|
| Factorial ANOVA | Metadiscourse corpus study (PLOS ONE) | 9/9 F-values exact to 2 dp |
| CFA + measurement invariance | Quality-of-life scale study (Sci Rep 2023) | 13/15 exact + 2 pattern |
| Mediation / moderation | Emoji leadership study (PLOS ONE) | 5/5 PROCESS estimates exact |
| + 15 more audits | GLMM 12/12 fixed effects · network science (exact LR) · ML classifiers 88–90% AUC · latent growth 20/24 · eye-tracking LMMs · VLM… | per-paper comparison tables |
Nineteen replication audits in total — each with a per-paper comparison table you can check yourself in 15–30 minutes. The audits span the fields you actually work in: psychology, linguistics, education, and business studies (management and marketing).
The same machinery runs on your own data — with a verification battery standing in for the published answer key. See what that looks like →
What I can build into your grant
Every pipeline is a working, reproducible demo — script, data, report — adaptable to your data in days. The first four are technical signal-processing pipelines (EEG, speech, eye-tracking) that no off-the-shelf AI tool can deliver; the rest are validated LLM methods with published reliability metrics. Prices are fixed per deliverable; the first one is free if not useful.
| If your research involves… | The pipeline I run | Validated result |
|---|---|---|
| EEG / ERP | Full ERP pipeline: filtering → epochs → components | N1 ~85 ms @ Fz, matched |
| Speech, phonetics, listening | ASR transcription + acoustic analysis (Praat-grade) | F0, duration, intelligibility |
| Child speech / speech disorders | Child-Cantonese ASR + error-preservation analysis | SSD errors tracked in carrier frames |
| Eye-tracking reading | Fixation/reading-time mixed models | all p < 1e-8, matched |
| Interviews, open-ended responses | LLM-assisted qualitative coding with reliability check | κ 0.81 / 93% agreement |
| Reviews & screening | Systematic-review screening with LLM | κ 0.88 vs manual screening |
| Surveys, scales, psychometrics | Reliability, factor structure, invariance testing | α .81–.85, 100% recovery |
| Experiments (RT) | Reaction-time analysis: outliers → stats → figures | d = 0.50, matched to paper |
| Corpus / text analysis | NLP features, sentiment, AI-text detection, classifiers | 97% accuracy / κ 0.93 |
| Business: consumer text, surveys, experiments | Sentiment & themes at scale, scale validation, moderation / mediation pipelines | 5/5 PROCESS estimates exact (marketing audit) |
| Humanities: archives, media, discourse | Corpus & discourse analysis, AI-text detection, digital methods | published discourse frame reproduced (61% vs 65%) |
| Translation / interpreting | Quality evaluation + terminology consistency | 100% preference, 90% error detection |
See every demo, with its honest limits stated →
No data yet? The pipeline starts before the data.
Most of the labour — and most of what can go wrong — happens before the first data point. These are working demos too, each adaptable to your project in days.
Literature & synthesis
A structured literature matrix from your reading list — design, sample, measures, findings per paper — plus screening with validated reliability and a synthesis of the gaps your proposal needs.
Evidence: 11 real papers turned into one matrix with a reliability check (κ 0.76–0.88 vs a deterministic baseline). See the demo →
Instruments & experiments
Bilingual questionnaires forward-translated, blind back-translated, and checked against the published Chinese version before you pilot — plus experiment materials, power analysis, and pre-registration drafted.
Evidence: GAD-7 / PHQ-9 items validated against the published Chinese; a known injected error was caught and repaired by the loop. See the demo →
Grant proposals
Your draft scored against the official RGC criteria before submission — where the panel will attack, and the patch — plus reviewer simulation and budget sanity checks.
Evidence: official GRF2 2026/27 criteria, scored criterion-by-criterion with verbatim evidence quotes. See the demo →
How it's paid for: these are research work like the analysis — funded from the same grant lines. And they compound: the literature matrix feeds the proposal, the proposal funds the data, the prepared pipeline analyses it.
See it. Try it on your data. Build for your project.
Three steps, each de-risked: watch the machinery on public data, test it on your own published numbers, then build for your current project.
1 · See it
A 15–30 minute walkthrough of the demos — working pipelines on public data, with their honest limits stated. Nothing to install, nothing hidden.
2 · Try it on your data
One trial deliverable on your own topic — your published data, checked item-by-item against your own paper's numbers. Fixed HK$3,000–8,000, free if not useful.
Start by email — no call needed. Get an indicative quote first →
3 · Build for your project
A full pipeline on your current data — milestone-based, delivered with documentation and handover, with ongoing support if you want it.
Not sure which step fits you? Answer five questions and find your scenario — 30 seconds →
Also on your desk: the rest of an academic's job
The same discipline — a working, verifiable pipeline, fixed fee, free-if-not-useful trial — applied to the other two parts of your job.
Admin agent
Grant applications, ethics/IRB paperwork, reference letters, meeting minutes, inbox triage — run by an AI agent with a human in the loop, tuned to how you work.
Evidence: working demo — inbox triage, meeting minutes, and reference-letter drafting, with human confirmation on every action.
How it's paid for: dean's discretionary / department / TDG budget. Two-week free pilot, then setup + admin retainer (HK$5–8K/mo — versus HK$15–30K for a real admin RA).
Teaching modernization
Your existing lecture decks, rebuilt: gap analysis against the current field, a rewrite that keeps your voice, redesigned slides with speaker notes throughout.
Evidence: a complete rebuild of a 2022 lecture deck — before/after you can view in minutes.
How it's paid for: TDG / department teaching budget — never GRF lines. One-deck trial, free if not useful; per-deck HK$2–5K, or a per-course package.
Tried ChatGPT? It forgets everything between chats.
That's not a failing on your part — it's how those tools are designed. I set up a private AI assistant that remembers — every conversation, correction, and decision — and does the work: it gathers data from the web, runs analyses, drafts documents, builds spreadsheets, reads PDFs, and writes code. Because it remembers, none of it starts from scratch — it knows your data, your methods, and what you approved before.
- Reference letters: you correct a draft once; it never repeats that mistake, and next semester's letter starts from what you approved, not a blank page.
- Analysis: you explain your pipeline once; next month, the same analysis on new data runs without re-explaining the steps, file formats, or codebook.
- Grant applications: it remembers your CV, your institution's requirements, and what worked in your last successful application.
You choose which models it uses — and for the most sensitive material it can run entirely on your own server, so nothing trains on your data and nothing leaves your control. The setup includes onboarding: a guided walkthrough and a plain-language guide, so you're not left with a black box.
How we work
Nothing is hidden. Every engagement runs the same five steps — in replication mode (published data, checked against the paper) or build mode (your own data, verified against the data itself). Read the full five-step account →
- 1
Know the data — public or yours
The deposit behind the paper, or the inventory of your own dataset — structure, codebook, provenance.
- 2
Load and sanity-check
Row and column counts, value ranges, and missingness appear before any analysis.
- 3
Build or rebuild the analysis
The paper's models rebuilt — or your analysis constructed — in an open statistical stack.
- 4
Verify — against the paper, or against the data itself
A side-by-side comparison table, or a verification battery for new data.
- 5
Report and hand over
Comparison table / verification report, plain-language summary, and the reproducible script and data.
Contact
15–30 minutes, no obligation — to see your own numbers reproduced, or to scope a trial deliverable on your topic. Prefer email? Get an indicative quote first, or send your paper or data link and start the trial without a call.
Replies within 24h on business days — same day for urgent.
Working arrangement
Video kickoff in week one. Fixed-fee pricing — no hourly meter, no overhead.