Research · Workflow
Research with AI, checked by people
I run evidence reviews and meta-analyses with AI doing the legwork. I choose the question, decide at every gate, and bring in human reviewers before anything goes out.
This page shows the whole process, from spotting the need to the final human check, with three anonymized projects: a water reuse meta-analysis, a psychiatric medication safety meta-analysis, and a patient article library for a clinic.
AI does
Searching, screening, data extraction, statistics, figures, drafting, and its own error checks.
I do
Pick the question, make the go or stop calls, get the papers AI cannot reach, and challenge the story.
Other people do
A statistician, a second data extractor, or a licensed clinician reviews before release.
Why
Six reasons to do research this way
01Answer a real question
A plant, a regulator or a patient has a decision to make, and the published evidence is scattered across dozens of papers and reports.
02Find out early if it is answerable
A feasibility check takes hours, not months. Most ideas should die here, cheaply.
03Do it alongside a day job
A review that needs a funded team becomes possible for one working engineer.
04Make every number traceable
Each value carries a source locator, and every table and figure regenerates from scripts.
05Build real expertise
Reading and pooling a whole field's data is the fastest way I know to become the reference person on it.
06Publish content people can trust
Public health articles get the same rule as papers: drafted fast, released only after an expert signs off.
The process
Eight stages, each with a gate
Tap a stage. Every one ends with a decision I make, and "stop" is always an option.
Explore and eliminate
Four ideas in, one paper out
This is the real path behind the water reuse paper. Three ideas were cut or turned before the fourth earned a full study.
Membrane stack geometry in electrodialysis
Hundreds of candidate records, but studies did not report comparable numbers. No pooled effect could be supported, so it shrank to an evidence map.
Normalized energy comparison across desalination designs
The novelty check showed earlier work had already done the core calculation. A new paper would need to add something those did not.
An industrial water-reuse question
Went through a formal go or no-go check. The original framing did not pass, so the question was redrawn instead of forced.
A treatment-credit question in potable reuse
Clear decision for regulators and utilities, enough independent studies, extractable numbers, and no existing review answering it.
Examples
Three projects, anonymized
Titles, facilities and trial names are left out. Neither paper is published yet, so findings stay private. The numbers below describe the process.
Quality control
What the humans do
AI checks its own work with tests and audits. That is necessary and not sufficient, so these steps stay with people.
A stage counts as done only with dated output evidence plus a recorded human review. Instructions alone never count.The completion rule built into my research engine
Take it
Use the same method
Paste this into any capable AI assistant and give it your subject area.
Starter prompt
You are running an evidence review for me. You do the work end to end. I decide at each gate and fetch any papers you cannot open. Never ask me to extract data by hand. A. From my subject area, list 5 to 8 questions a pooled analysis could answer. For each: is there a recent review already, how many studies report extractable numbers, are outcomes comparable, and who would use the answer. Drop any with fewer than about 5 usable studies. Show me a ranked table and stop. B. For the question I pick: write the one-sentence headline the paper would lead with, what result would kill it, and the 3 to 5 closest existing papers. Extract the 5 easiest studies and run a rough estimate with uncertainty. Tell me honestly what kind of paper this can be. Stop. C. Fix the protocol before extraction: eligibility, model, missing-data rules. Search at least two databases plus citations and agency or utility reports. Log every query. D. Extract one row per reported value with a source locator. Keep unique record IDs. Flag censored values. E. Build the analysis so every number, table and figure regenerates from scripts. Stress-test each claim the day you add it. After the first full run, restate the headline using only what survived. If it weakened, tell me at once. F. Write findings first, in plain language. Do not let the wording claim more than the data. Declare AI use. G. Prepare a review package for a human statistician and a second extractor. List known weak points up front. At every gate report: what holds, what changed, what it means for the headline, and your recommendation, including "not worth pursuing" when that is true.
Have a question the literature should answer?
I take on projects: I develop the question, run the study and handle the quality control. If you would rather do it yourself, the starter prompt above will get you going.
shahruuz@gmail.com
LinkedIn