Behind Grace, a conversation
How we built Grace
A finger-prick test, a positive result, and six weeks of waiting for the venous draw that would confirm or rule out type 1 diabetes in our son. Out of those six weeks came a single observation, every family who walks into a clinic wants to learn at their own pace, and most of the time the clinic cannot move at theirs. Seven years on, that observation became Grace.
Ask Grace
Want to feel how Grace works before you read about how she was built? Ask her one question on a topic you already know well, and see what comes back.
Why this page exists
Most people who arrive at Grace ask the same first question, in different words: who built this, and what is the philosophy underneath it? The honest answer takes more than a paragraph, because Grace was not designed in a sprint. She is the result of seven years of writing diabetes guides at the kitchen table, two years of evidence ingestion, three months of architecture rebuilds, and one long conversation with a friend who happens to be a senior AI developer.
In May 2026, Arseniy Arsentyev invited John on his podcast Tomorrow’s Medicine to walk through the build. This page is a decoded transcript, organised by the questions people ask us most, with timestamps that jump straight to the moment in the recording where John answers each one.
Watch the conversation
The build, decoded
Fourteen moments in the conversation that answer the questions we hear most. Each disclosure opens to a short summary; the “jump to this moment” link opens the recording on YouTube at the second John starts speaking on the topic.
The six-week wait, and why Grace exists at all
John’s son had a positive antibody result on a finger-prick screen. The venous confirmation that finally returned negative took six weeks. In those six weeks, sitting at the dinner table watching his son eat, John realised that all the dietetic theory in the world meets a different kind of difficulty when the child in front of you is the one running the experiment. That is the moment Grace started.
The 80, 20 line, in John’s own voice
The strapline that anchors Grace, gets you 80 per cent of the way there with 20 per cent of the effort, the final 20 per cent takes self-discovery, guided by human expertise. John explains why the 80 is a feature, not a limit, and why the 20 is on you and your network.
Type 1 and Type 2 diabetes, explained in one breath
A three-minute explainer that Arseniy stops the conversation to praise. Insulin as the key, the cell receptor as the lock, autoimmune destruction versus insulin resistance, and why one is reversible early and the other is not.
The three layers, base, delivery, philosophy
Grace is not a chatbot wrapped around an LLM. Layer one is the consensus guidelines (ISPAD for paediatrics, ADA and EASD for adults). Layer two is seven years of translation work, taking dry guideline language and reshaping it for a ten-year-old or a seventy-year-old grandparent. Layer three is the philosophy, which is the one most builders skip.
Grade A to D evidence, why D is not zero
Every educational recommendation Grace makes carries an evidence grade. A is systematic review or randomised control trial replicated across studies. B is a single controlled trial. C is large observational data, the registries that tell you what happens in the real world, which is often more informative than the trials. D is expert consensus, useful, sometimes the only thing available, and worth flagging as such.
Exercise, three levers not fifteen
Ask any person with type 1 what affects glucose during exercise and they will name fifteen variables. Ask them which three matter most and the conversation slows. John names them: where your glucose is right now, whether it is trending up or down, and how much insulin is still on board from your last dose. Three answers, a plan you can actually follow.
Why Grace does not remember you, on purpose
Grace has no memory of previous conversations. That is a design decision, not a limitation. Memory implies a relationship, and a relationship implies the right to make recommendations specific to you. John explains why he believes that right belongs to a human who can see you, and why an educational advisor that pretends to know you is more dangerous than one that openly does not.
Sponsorship, educational grants, and the five questions
GNL takes educational grants but never sponsorship. The difference matters, an educational grant is signed money to create content; a sponsorship is rented attention. Grace will only educate on devices that meet a five-question evidence bar on accuracy studies. Of fourteen CGMs available in Europe, five meet it. The other nine do not appear.
Building Grace was almost a bankruptcy event
The first architecture was flat, everything in one big context for the model to wade through. Cost per query came in higher than anticipated. Two thousand pounds disappeared in two weeks before the structure changed. The fix, a wiki layer with retrieval over a database, cut the same workload to roughly five per cent of the cost. Without that change Grace would not have made it to launch.
The plus-one model, in plain English
Grace is free for people living with type 1 diabetes. If you find her useful and want to pay it forward, a five-pound contribution covers your access and funds one additional account for someone on the waiting list. A manufacturer or charity that sponsors a thousand accounts unlocks another thousand. The door revolves; the access stays free for the people who need it most.
Grace Max, used to write a manuscript
John sent his academic mentor a paper Grace had drafted at 80 per cent, polished by him on the final 20. Her reply, this is better than I could have written, how did you do this? The point of Grace Max is not to remove the writer from the work; it is to remove the dog-crap-first-draft from the writer’s evening.
Agents for manufacturers, only for products we believe in
Three diabetes-adjacent companies have asked Grace’s team to help build their internal agents (marketing, compliance, education, sales, ethics) on top of the same grounded source of truth. John’s filter is single-sentence, we will help you if your product is good for people. Continuous lactate monitoring, yes. Continuous ketone monitoring, not yet. Continuous glucose monitoring as a wellness gadget, no, the signal-to-noise ratio is too thin.
Lactate, the canary in the coal mine
Continuous lactate monitoring, in John’s view, is the technology that will identify metabolic disease ten to fifteen years before glucose does. Not a Type 1 tool, an everybody tool. By 2028 to 2030, insurance companies will be paying for it. The Type 1 community gets to watch this wave land first because we have been wearing biosensors for a decade.
Eighty per cent of jobs change, communicators thrive
The closing third of the conversation turns to the wider question, what does AI do to a profession? John’s answer is a Taleb-shaped one, the people who can adapt fast and communicate well are fine; the people who hide behind procedure are not. Doctors, nurses, dietitians and teachers who can look someone in the eye and read what is not being said are the currency of the next decade.
The three Taleb filters underneath Grace
Most of the build choices that look unusual from the outside trace back to three ideas from Nassim Nicholas Taleb. The reason they show up in a diabetes education tool is that diabetes management lives in the same neighbourhood as the problems Taleb writes about, rare events with permanent consequences, asymmetric information, and the temptation to optimise for a metric instead of a person.
Black Swan, precaution before novelty
Type 1 diabetes is a Black Swan condition. Things can go right for a decade and one bad night undoes the whole run. Grace educates on devices that have cleared a five-question evidence bar; the rest she leaves alone. Not because the rest is bad, but because the risk of false confidence outweighs the convenience of inclusion.
Via Negativa, clarity by subtraction
Most of the work in building Grace was not writing more, it was cutting more. Out of a thousand papers each week, Grace’s appraisal layer surfaces ten; John selects five. The signal compounds. The output, when you ask Grace anything, is the three levers that matter, not the fifteen that flatter the model.
Skin in the Game, no prescription, no memory
Grace will not tell you what to do, will not remember who you are, and will not pretend to know your context. That is not a feature backlog; it is the design. The 20 per cent that Grace cannot reach is the part that should belong to a clinician, a peer, or a family member who can see you.
The single trade-off. Grace is intentionally less personal than she could be. The cost is that she will sometimes feel generic. The benefit is that you can trust the framing, because nothing about her output is being optimised for you specifically, and that is what allows you to take it to someone who knows you and ask the next question well.
What this means for you
For people living with type 1 diabetes and their families
Use Grace the way you would use a well-read friend who happens to be a clinician. Bring her your specific question, take what comes back as a starting point, and then take the starting point to the human who knows your life. The 80 per cent is hers; the 20 per cent is the conversation she could never have.
- Pick one question you already think you know the answer to, and ask Grace anyway. The reply will tell you whether her register fits your style.
- If a Grace answer carries a Grade C or D label, that is not a flaw, it is the most useful warning she can give you. Bring those answers to your team first.
For clinicians and educators
The threat in this technology is not that it replaces the bedside conversation; the threat is that it makes the consultations of clinicians who hide behind procedure look thinner than they used to. The opportunity is the opposite, you spend less of the appointment getting someone to 80 per cent and more of it on the part that you, specifically, do well.
- If you find yourself spending an appointment relaying material that Grace could have delivered first, consider whether the bottleneck is the material or the format.
- Grace cites the evidence grade on every answer; that is a discussion prompt for your team meeting, not a number to argue with.
For manufacturers and partners
If your product clears the evidence bar, we will work with you. If it does not, we will tell you what would clear it. We do not take sponsorship; we accept educational grants on signed terms, and your product makes it in if the data supports it whether you fund the work or not. The relationship is faster when there is a relationship, slower when there is not. That is the whole influence model.
- The pathway is on our consultancy page, not on Grace’s surface; she stays a clean educational product.
- The five-question CGM evidence bar is the same one we ask of every CGM, including the ones whose makers have funded the work.
About the conversation
Tomorrow’s Medicine is a podcast hosted by Arseniy Arsentyev, a computer scientist who interviews builders working at the edge of healthcare. Arseniy reached out to GNL cold after coming across the site, and was the one who pushed the conversation past the surface-level build story into the philosophy and the economics underneath. The questions he asks midway, the bond-rating analogy for the evidence grades, the philosophical pushback on educational grants, the language-model-as-correlation-engine point near the close, are the ones that make this episode different from the usual founder interview. Episode 14 was recorded in May 2026.
John Pemberton is a paediatric diabetes dietitian at Birmingham Children’s Hospital and the founder of The Glucose Never Lies. He has been writing diabetes education for ten years and living with type 1 diabetes for fifteen.
Where to go next on GNL
Behind Grace
How we built Grace
This page is for educational exploration only. It describes the build of Grace and the philosophy behind it. It is not medical advice and cannot replace individual clinical guidance from your diabetes care team.