Home The short version The work Six challenges, eleven systems Background What I draw on Work with me Get in touch
Home  /  The work

Six real-world challenges, and what I built for each.

Naming what is wrong with generative AI is the easy half. Each challenge below is structural rather than a bug awaiting the next model release — and each is followed straight away by the thing I built in response, so you can judge the diagnosis and the answer together. Eight of the eleven systems below are publicly available today; the last three are everyday productivity tools, kept separate because they answer no structural challenge at all.

11Systems builtScoped, engineered, shipped and maintained solo
8Publicly availableOpen to anyone, with no institutional login
2Browser extensionsInstallable worldwide, including from the Chrome Web Store
3Never leave your browserThe task list, timer and calculator contact no server at all

The portfolio, positioned

PROTECTS INTEGRITY & JUDGEMENT REMOVES DRUDGERY WITHIN ACADEMIC PRACTICE BEYOND ACADEMIA RAG Study Screener Full-Text Screener RIS Deduplicator Journal Intelligence Scholar Metrics ABDC Ratings Evidence4Impact ELT Pub Private To-do List Research Timer Calculator Live and public Self-hosted or research build
Position is editorial, not measured. It is there to show the shape of the portfolio: the work spans the full range from removing mechanical drudgery to protecting judgement, and it does not stop at the edge of the university.
Challenge 01

Generating became free. Checking did not.

The cost of producing a plausible draft, a summary, a literature overview or a set of screening decisions has collapsed. The cost of establishing that any one of them is correct has barely moved — it still requires a qualified human reading a primary source.

Every hour an organisation saves in production is quietly borrowed against an hour of appraisal that nobody has scheduled. Across a team and a year, that borrowing compounds into assurance debt: material in circulation, cited and acted upon, that no competent person has actually verified.

The widening gap

Cost to verify one item Cost to generate one item ASSURANCE DEBT Volume of AI-assisted output → Cost per item ↑
Schematic, not measured data. The shape is the point: the two curves move in opposite directions, and the shaded area is work that someone is assumed to have done.
What I built for it

Screening tools that make checking cheaper than redoing. Each decision comes back with the passage that justifies it, so a second reviewer audits reasoning rather than re-reading four thousand abstracts — and the reported screening counts become reproducible instead of approximate.

01Research build

RAG Study Screener

Screens records against your own protocol and returns an IN / OUT / UNCLEAR verdict with the passage that justifies each call. Model-agnostic across Anthropic, OpenAI and Google.

ForReview teams, evidence units in government and NGOs, research groups
WhereSelf-hosted build · open to collaboration
Available for collaboration
02Research build

Full-Text Screener

Reads past the abstract and attaches the supporting text to every eligibility decision, turning an exclusion log into evidence rather than recollection.

ForSystematic reviewers, doctoral candidates, methodologists
WhereSelf-hosted build · in trial
Available for collaboration
03Shipped

RIS Deduplicator

Merges every database export, matches on DOI, title and author–year, and logs exactly what was removed — so the PRISMA count you report is reproducible.

ForResearch librarians, review teams, information specialists
WhereWeb and desktop builds · runs on your own machine
Self-hosted — ask for a build
Challenge 02

Confident output, and no way back to the source.

A fluent answer carries no signal about its own reliability. A fabricated reference, a misattributed finding and a quietly dropped qualifier all arrive in exactly the same calm register as a correct statement. The register is the failure mode: it is tuned for plausibility, and plausibility is what an expert uses as a first-pass filter.

You cannot audit what does not record where it read. When a claim is challenged six months later — by a journal editor, an auditor, a replication attempt, or a court — “the model said so” is not a defensible answer.

Two ways to reach the same decision

Default route Your question ? A confident answer NO SOURCES RECORDED How the tools do it Your protocol + the records DEFINED CORPUS RETRIEVAL CITED PASSAGES IN OUT UNCLEAR “…all participants were undergraduates, so criterion 3 is not met.” A HUMAN MAKES THE CALL
The difference is not accuracy on a good day. It is whether the decision can still be defended on a bad one.
What I built for it

Retrieval-grounded appraisal: the tool reads a defined corpus, returns an IN / OUT / UNCLEAR verdict, and attaches the supporting text to every decision. The exclusion log becomes evidence rather than recollection. Runs on Anthropic, OpenAI or Google models — the model is a component, not an oracle.

04Live

Evidence4Impact

Ask a question in plain language. It writes and shows the search string for each database, searches PubMed, Europe PMC and OpenAlex, ranks hits by study design from the records’ own publication types rather than a model’s opinion, and answers with every statement linked to its source — checking each citation exists and flagging wording that claims more than the studies support.

ForPractitioners, policy officers, clinicians, educators and researchers who need an answer today
WhereLive at evidence4impact.org · free to try, accounts for higher limits
Run a search ↗
Challenge 03

Integrity is also a question of whose evidence gets in.

Generative AI did lower the cost of producing a submission, and low-governance venues are built for exactly that gap. But treating research integrity as a predator-detection problem misses most of it. Editorial practice itself decides what enters the record — which topics count as being of general interest, which affiliations and which varieties of English read as authoritative, which author lists receive a sympathetic desk review rather than a swift rejection. Those decisions are rarely malicious and they compound regardless.

What results is a literature whose coverage is uneven in ways its readers cannot see. Sound evidence from under-represented regions, languages, institutions and career stages surfaces late or never, while the metrics and lists we reach for as proxies of quality largely encode the hierarchy that produced them. An early-career researcher weighing an unfamiliar invitation is making two judgements at once — is this venue legitimate, and is this system fair to someone like me — and has almost no help with either.

Who gets into the record

WHAT IS SUBMITTED WHAT REACHES THE RECORD EDITORIAL PRACTICE DESK REVIEW · REVIEWER POOL LANGUAGE & FORMAT NORMS The record is not a representative sample of what was submitted — and nothing in it shows the reader the difference. Author groups — region, institution, language, career stage
Schematic. The proportions are not a measurement; the asymmetry is the point, and so is the fact that it is invisible from the published side.
What I built for it

Integrity signal at the moment of the decision, and more than one kind of it. Journal Intelligence profiles a venue from publication growth, title warning lists, and authorship concentration — which is an equity signal as much as a fraud one, since it shows when a journal publishes disproportionately from a single group. The Scholar Metrics and ABDC overlays put quartiles, indexing and tiers in front of you while you triage, so the hierarchy is at least visible and arguable rather than silently assumed. None of them tells you a venue is fair; they tell you where to look.

05Live

Journal Intelligence

An integrity-risk profile for any journal, assembled from publication growth, authorship concentration and title warning lists — in the time it takes to paste an ISSN.

ForEarly-career researchers, research offices, librarians, editors
WhereFree web tool · no account
Open the tool ↗
06Live

Scholar Metrics

Scopus quartiles, SCImago indicators and Web of Science indexing badges rendered inline on Google Scholar results, while you are still triaging.

ForAnyone who reads Google Scholar — researchers, students, practitioners
WherePublished on the Chrome Web Store · free
Install the extension ↗
07Live

ABDC Journal Ratings

A*, A, B and C tiers from the ABDC Journal Quality List overlaid on Google Scholar, with a tally of each tier at the top of the page.

ForBusiness and management academics, business schools, doctoral students
WhereBrowser extension · distributed via ELT Pub
Get the extension ↗
Challenge 04

There is still no standard way to say what the AI did.

AI-assisted work is entering the permanent record now, and there is no shared convention for declaring how. “AI was used in the preparation of this manuscript” covers both a grammar check and a fabricated dataset. Readers cannot distinguish between them, so they discount all of it — or none of it.

Underneath sits an older problem that AI has made acute: a single label, peer reviewed, is asked to carry editorial approval, methodological scrutiny, and trustworthiness all at once, with no indication of who did what, on what basis, or with what conflicts. One word cannot hold that much weight.

Pulling one label apart

One opaque label peer reviewed BY WHOM? ON WHAT BASIS? Carries approval, scrutiny and trustworthiness in one word — and says nothing about AI use. Three attributed claims Editorial approval NAMED EDITOR · DATE · VENUE POLICY Independent verification EXPERTISE · BASIS · DECLARED CONFLICTS AI-use disclosure WHICH STEP · WHICH MODEL · WHO CHECKED PERSISTENT ID · NOT VENDOR-OWNED A reader can now tell an approved record from a scrutinised one.
Separating the claims costs nothing and resolves most of the ambiguity. The reason it has not happened is not technical.
What I do about it

The standard does not exist yet, and I argue for one: approval, verification and AI-use disclosure stated separately and separately attributed. Evidence4Impact is what that looks like implemented in a single tool — it publishes every search string so the search can be repeated, ranks by study design while saying plainly that design is not the same as how well a study was conducted, verifies that each citation it hands you exists, and flags its own wording when it claims more than the studies support.

Challenge 05

What you paste in does not come back out.

Professional work is full of material that should not leave the building: participant data, unpublished study titles, embargoed dates, client records, draft findings, student work. The default route to any AI-assisted convenience is an account, a vendor’s servers, an unfamiliar jurisdiction, a chain of subprocessors, and a privacy policy that can be revised without your involvement.

The quiet scandal is how little of that is necessary. A countdown timer, a task list, a calculator, a deduplication pass — none of these require a data relationship with anyone. They acquired one because collecting data became the default business model, not because the task demanded it.

Two paths for the same keystroke

YOUR SENSITIVE WORKING DATA The default path ACCOUNT VENDOR SERVERS RETENTION WINDOW SUBPROCESSORS JURISDICTION POLICY MAY CHANGE Five parties you did not choose, and one you cannot audit. How these tools do it YOUR SENSITIVE WORKING DATA Your own browser storage NO ACCOUNT IS POSSIBLE Full stop. NOTHING IS TRANSMITTED
Most everyday research utilities need no network at all. Where the work genuinely needs a model, the build is yours to host.
What I built for it

One rule, applied to everything I build: if a task does not need a server, it does not get one. The heavier research tools ship as self-hosted builds, so sensitive corpora never leave infrastructure you control. The small everyday utilities — the three further down this page — hold their data entirely in your own browser, where no account is offered and none is possible. Neither is a feature anyone asked for; both are what the work actually required.

Challenge 06

Judgement is being delegated by accident.

Nobody sits down and decides to stop thinking. It happens one default at a time: accept the summary, accept the suggested framing, accept the ranking, accept the draft conclusion because it reads well and the deadline is Thursday. Each individual acceptance is defensible. The aggregate is not.

The risk here is not a dramatic error. It is a slow transfer of exactly the parts of the work that constitute expertise — noticing the anomalous case, distrusting the tidy result, knowing which qualifier matters. Those are not tasks that announce themselves when they are skipped.

What may be delegated, and what may not

Merging exports Formatting references First-pass screening Drafting, summarising ASSIST, THEN CHECK Eligibility calls Interpreting a finding Accountability for the claim Mechanical Constitutive of expertise EVERY TOOL HERE STOPS AT THIS LINE
The line is a design decision, taken once, before the first feature. It is why none of these tools offers to write your conclusion for you.
What I do about it

Research on Human–AI Distributed Critical Thinking: where the cognitive work can legitimately be shared, and where sharing it hollows out the practitioner. In build terms it is a hard design rule — every tool stops one step short of the call, states its limits in the interface, and makes the human decision point explicit rather than quietly absorbing it. The independent writing at ELT Pub argues the case for people with no vendor telling them otherwise.

08Live

ELT Pub

An independent, scholar-led publication and toolmaker — founded and run solo — turning research on generative AI in education into explainers, critical commentary and the free utilities listed here.

ForTeachers, teacher educators, language centres, school leaders
WherePublic website · free to read, nothing to sell
Visit the site ↗
Beyond the research work

Everyday productivity tools.

Not everything I build answers a structural challenge. These three are ordinary working utilities — a task list, a countdown, a calculator — and they are here because they are held to the same rule as everything above: useful without becoming a data relationship. No accounts, no telemetry, nothing transmitted. They are small, and the point they make is not.

01Live

Private To-do List

Dated tasks, priorities and progress for work that holds participant details and embargoed titles — kept entirely in your own browser, where no account is offered and none is possible.

ForResearchers, clinicians, consultants — anyone with confidential tasks
WhereFree web tool · nothing is transmitted
Open the tool ↗
02Live

Research Timer

An analogue countdown readable across a room, for observations, think-alouds and timed protocols, with an independent stopwatch. Deliberately unquantified: it keeps no record of you.

ForClassroom teachers, facilitators, researchers running protocols
WhereFree web tool · works offline
Open the tool ↗
03Live

Calculator

A fast calculator with a running tape and full keyboard support, so you can see how you reached the number. No logged queries. Not everything needs to be intelligent.

ForAnyone, inside a university or well outside one
WhereFree web tool · opens instantly, works offline
Open the tool ↗

Automation assists.
People decide.

Every tool listed on this page is a screening, reflection or workflow aid. Outputs are meant to be checked against primary sources, disciplinary standards, local context and informed human expertise. None of them is a substitute for the judgement of the person accountable for the decision.

This is not a disclaimer bolted onto the end. It is the constraint the tools were built under — which is why they show their reasoning, name their own limits, and stop short of the call that belongs to you.

Where the work lands

HOW IT IS DISTRIBUTED WHO IT REACHES Chrome Web Store2 EXTENSIONS · PUBLIC Free web tools4 TOOLS · NO ACCOUNT Self-hosted buildsRUNS ON YOUR INFRASTRUCTURE Platform & publicationELT PUB · EVIDENCE4IMPACT Researchers & reviewers Students & early-career researchers Business & management academics Publishers & integrity offices Teachers & educators Independent educators Professionals with sensitive data Within academia Beyond it
Three of the four channels require nothing of the person using them — no institution, no subscription, no account. That is deliberate: it is what lets work built inside a university be used by someone who has never set foot in one.
Shipped, not proposedEvery item here exists and runs. Two are research builds still in trial, and they are labelled as such.
Free and account-freeNothing asks for a sign-up, and most of it never contacts a server at all.
Limits stated in the interfaceEach tool names what it cannot tell you, at the point where you are relying on it.
Yours to hostWhere the data is sensitive, the build runs on your infrastructure and the record stays with you.

Start with the problem, not the technology.

There is nothing to prepare and no process to go through. Tell me what you are facing, in a few sentences, and we can work out from there whether I am any use to you.

Work with me