The Problem: An ML Researcher Meets a Non-Technical Panel
Dr. Aisha Kaur has spent five years building machine-learning systems that turn messy, real-world GP consultation audio into structured clinical notes. Her model - trained on 12,000 de-identified transcripts - handles ambient noise, regional accents, and interruptions better than any off-the-shelf solution. She’s published, she’s patented, and she knows the transformer-encoder-decoder literature inside out. So when she spots an Innovate UK Smart Grant calling for disruptive AI in healthcare, she thinks: this is mine.
Then she reads the assessor criteria.
Innovate UK panels aren’t staffed by AI researchers. They’re industry practitioners, commercial directors, and NHS procurement leads - people who judge projects on viability, delivery risk, and market readiness, not attention-head depth or RLHF convergence curves. UKRI’s own guidance is blunt: “Your application is likely to be seen by many people, including some who will not be familiar with your particular specialism.”
Aisha doesn’t adjust her language. Her first draft opens:
“We employ a transformer-based encoder-decoder architecture with cross-attention over multi-modal embeddings (audio + ASR outputs), fine-tuned via reinforcement learning from human feedback (RLHF) to optimise clinical entity coherence and ICD-11 coding alignment.”
She hits submit. The feedback arrives six weeks later. One assessor comment stops her cold:
“We couldn’t understand what you’re actually doing - or how you’ll deliver it. We couldn’t score it.”
The panel didn’t doubt her expertise. They couldn’t see it. And because they couldn’t see it, they couldn’t fund it.
Over 50% of unsuccessful UKRI applications fail on poor communication of how, not what. Aisha’s proposal was technically brilliant - and functionally invisible. To win next time, she’d need to translate her machine-learning project into a language her assessors actually spoke: logistics, evidence, and human outcomes.
The Pivot: Translating Architecture into Action
Aisha stared at her draft for an hour before she saw the pattern. Every sentence was designed to prove she understood transformers, cross-attention, and RLHF. The problem? None of that proved she could deliver a working product in a GP surgery.
She deleted the architecture paragraph. Then she wrote one sentence that changed everything.
Our AI acts like an experienced medical secretary - listening to real GP consultations (ambient noise, accents, interruptions included) and drafting first-pass clinical notes in real time.
That was the pivot. Not dumbing down. Translating.
Map every technical feature to a human outcome
Aisha went through her model piece by piece. Each component got a new job description:
- Audio pre-processing → “Handles GP consultations recorded on standard room mics, not studio equipment.”
- Entity extraction → “Automatically pulls diagnosis codes, medications, and referral details into structured fields.”
- Confidence scoring → “Flags uncertain entries in yellow for clinician review - no silent errors.”
No more ‘encoder-decoder’ or ‘embedding alignment’. Just: what does it do, for whom, and when?
Anchor everything in concrete logistics
The assessor didn't care about Aisha's training loss curves. They cared about whether the thing worked outside a Jupyter notebook.
Aisha added three sentences of pure feasibility:
In a 4-week pilot across two NHS GP practices, doctors edited and signed 94% of AI-generated notes - saving 62% of documentation time per patient. We now need £680k to integrate securely with SystmOne and EMIS, test across 8 additional practices, and finalise GDPR/NHS Digital compliance - all within 18 months.
She listed licensing status (closed beta with one NHS trust), a contractor quote for EMIS integration (£47k), and a risk mitigation for dialect bias (stratified validation across 12 UK regional speech groups).
The goal shifted. Aisha stopped trying to impress with complexity. She started showing safe, credible delivery. The assessor might not understand attention heads. But they understood 94% edit acceptance, 62% time savings, and a clear path to deployment.
That was the whole point.
The Feasibility Anchor: Proving Delivery Without Technical Detail
Aisha’s next move wasn’t technical. It was logistical.
She sat down with her pilot data from the two GP practices that had tested her system. Four weeks. 62% reduction in documentation time per patient. 94% of AI-generated notes accepted with minor edits. Those numbers said more than any architecture diagram ever could.
She built a feasibility anchor - a set of concrete, verifiable proofs that her project would work, not just in theory, but in an actual NHS clinic with real patients, real interruptions, and real GMs who hate clunky software.
What she included
- Licensing and integration quotes: Contractor rates for SystmOne and EMIS integration - £48k, confirmed by a certified NHS supplier. No guesswork.
- Named risks with named fixes: Latency during peak EMR load is mitigated with parallel processing. Regional accent bias is handled via stratified validation across 12 UK speech groups.
- A timeline, not a wish: 8 additional practices, 18 months, £680k - broken into quarterly milestones with go/no-go gates.
- TRL alignment: By checking the Innovate UK Smart Grants eligibility window, she confirmed her system sat at TRL 4-5 (prototype validated in relevant environment). This single step saved her from automatic rejection - the #1 killer for first-time applicants.
“The assessor doesn’t need to understand your neural network. They need to believe you can build it, test it, and deploy it on budget.”
The effect on the assessor
When the panel reader saw “4-week live pilot, 2 practices, 94% acceptance”, they stopped wondering if the AI would work. They started asking whether the business model held up. That shift - from technical risk to delivery risk - is exactly where you want them.
Aisha’s grant now proved her project was low-risk, high-feasibility - and that’s a story any assessor can follow.
Key takeaway: Replace architectural depth with execution evidence. If you can show it worked somewhere real, for real people, with real constraints, you’ve already won the feasibility argument.
The Lay Summary Rewrite: The 300-Word Test
First Pass: Still Speaking Machine
Aisha opened the lay summary box and typed the same draft she'd just rewritten for the feasibility section. "A transformer-based encoder-decoder with cross-attention over multi-modal embeddings…" She stopped. This wasn't a lay summary. It was a confession of technical obsession.
She deleted it.
Second Pass: Vague Enough to Be Anybody
Round two was an overcorrection. "Our AI helps doctors write notes faster." True. Also true of speech-to-text software from 2016. Nothing in that sentence said machine learning, de-identified training data, or 94% acceptance rate. An assessor reading this would have no idea what was novel - or why it deserved £680k.
She deleted that too.
Third Pass: The Café Test
Aisha imagined explaining her project to a stranger at a café. No slides. No jargon. Just two minutes before their coffee arrived. She wrote:
"GP consultations generate thousands of clinical notes that doctors hate writing. Our AI listens like a trained medical secretary - capturing real conversations with interruptions, accents, and background noise - then drafts the notes in seconds. In a four-week NHS pilot, 94% of AI-generated notes were accepted by GPs with minor edits, saving 62% of documentation time per patient. We now need £680k to integrate across eight more practices and complete NHS Digital compliance within 18 months."
She read it aloud. Her dad could understand it. Her former PhD supervisor wouldn't find it dumbed down. The assessor would get the problem, the analogy, the proven outcome, and the next step - in 110 words.
Why the Third Pass Won
The formula was accidental but repeatable: Problem → Analogy → Proven Outcome → Next Step. No architectural claims. No implied expertise. Just a story the assessor could retell in the scoring meeting.
Aisha saved the summary and moved to the next section. For the first time, she believed her application might actually be read - and understood.
The Submission and What It Taught Her
Aisha hit "Submit" on a Friday afternoon. Six weeks later, the email arrived: Your application has been scored high enough to proceed to interview.
She read it three times. Not because she didn't believe it, but because she wanted to hold onto the feeling of being understood. The assessor - a serial entrepreneur from the medtech sector, not an AI researcher - had written: "This project clearly knows what it's building, for whom, and at what risk. The pilot data is compelling. The mitigations are specific. I can see how this gets delivered."
Not a single comment about her transformer architecture. Not one.
What Changed: From Expert to Leader
Aisha reflected on the shift inside her own head. Early drafts had been written to impress. She'd packed sentences with cross-attention and RLHF, hoping the assessor would see her as a legitimate AI scientist. But that instinct was backwards. The assessor didn't need to certify her technical chops - the university already had. What the assessor needed was confidence that Aisha could run a project.
Once she stopped trying to prove she knew machine learning and started proving she knew how to integrate it into NHS workflows, the writing fell into place. The team, the timeline, the eight additional practice sites, the £680k budget broken down by workstream - that was the evidence that mattered.
Her Translation Checklist
She keeps a printed copy pinned above her desk. You can use it too:
- Lead with the outcome, not the method. What changes for the user?
- Anchor every technical claim in a concrete number - hours saved, accuracy rate, pilot sites done.
- Replace algorithm names with analogies ("like an experienced medical secretary").
- Name the specific risks and exactly how you'll fix them (latency? parallel processing).
- Show you've already started - no one funds a fantasy. Validation data, contractor quotes, confirmed partnerships.
- Budget to match the story - a detailed spend says "I've thought this through" more than any architecture diagram.
Aisha still uses those six rules every time she opens a grant application. They turned her from an ML specialist into a credible project leader - and made the assessor her ally, not her examiner.
Grant application writing isn't about proving you're smart. It's about proving you're ready.
Draft the next application in GrantGunner
Create a free account, add your documents, and draft the form with your team. Create a free account