I build the product and the operation around it.

Students without a tutor at home get stuck alone, and a general chatbot just gives them the answer.

• Coaching, practice, marking and assignments as separate agents, each graded by its own evaluator, across Maths, Physics, Chemistry and Biology for AQA, Edexcel, OCR and IB.
• Red-teamed the failure that matters most: a student talking the tutor into handing over the answer. Then a pass/fail judge on every session.
• A four-agent fleet read every production trace overnight, replaced a planned ops hire and produced half the evidence in the company retro.

30 → 850+

schools, from the beta to the national programme

67% → 99%+

marking on the golden set: bare model, then the syllabus we covered

33% → 65%

Socratic pass rate, doubled in two weeks, graded on every session

50,000+

tutoring sessions

Harris Federation's AI lead told us admin visibility and safeguarding were gate conditions, not features. Until both were met, the question was whether a school would let the tutor near students at all.

• Safeguarding detection shipped as the first feature, tuned to over-flag, with the designated safeguarding lead holding every judgement.
• Tuned it loud on purpose, then split conduct from safeguarding so the lead only gets paged for what matters.
• Built to the DfE's generative AI product safety standards for under-18s: session cut-offs, never presenting as human, and concerns escalated to a named person.

97%

fewer false safeguarding alarms, with no disclosure missed

2nd of 8

in the DfE's AI Tutoring Tools Pioneers Programme, picked from 53 bids

450,000

disadvantaged pupils a year the programme can reach

Teachers set one homework for a whole class and only find out who was stuck when they mark it. Help a student has to go and find mostly goes unused.

• Harris told us that to enter the curriculum we had to build for teachers, so the teacher became the way in.
• Homework built from the teacher's own material, after a head of maths said typing mark schemes was too slow.
• A weekly summary per class, safeguarding flags to named staff only, and the DPIA written before a data lead asks for it.

91% vs 23%

activation when teachers lead it, against students signing up alone

850+

UK schools

A document Q&A app for property lawyers showed "3 sources cited" under an answer about asbestos, in a lease that never mentions asbestos. The count was a regex over the model's own reply.

• Every claim carries a marker naming its quote, and code finds that quote word for word in the lease before it counts as evidence.
• A match jumps to the highlighted clause. A quote it can't find is drawn dashed, against the claim it weakens.
• Ruled out confidence scores and "stated vs inferred" labels: both are the model grading itself.

Code checks

every quoted passage, found word for word in the lease

Lawyer decides

whether the wording supports the point

Teams survey AI personas, then interview one to find out why it answered as it did. A plain chat renders fluent invention and grounded fact the same way.

• Each claim is solid if it traces to the respondent's survey answer or profile, and dashed if the model is reasoning past them.
• Asked to speak for people who weren't surveyed, the persona says it would be guessing.
• Ruled out a blanket "all views are fictional" notice. It concedes the problem and manages none of it.

Code checks

every claim against the respondent's own answer or profile

Researcher decides

what to quote, seeing how much rested on data

A home visit runs 40 minutes over, and the rest of a district nurse's list no longer fits her shift. Someone has to decide who is seen, who waits, and who owns the ones who wait.

• The assistant shows the shortfall and checks competency, care windows and shift time before a move reaches a nurse's screen.
• Ruled out letting the assistant propose that a patient waits. It flags what it can't place instead.
• Any visit that waits lands on tomorrow's list with an owner and a deadline.

Code checks

competency, care windows and shift time

Nurse decides

every clinical change, recorded with a reason

Selling into banks and law firms meant passing their due diligence before any programme could start.

• Built each programme into the partner's early-careers hiring, with HSBC, Morgan Stanley, KPMG, Ashurst, Accenture and Snap, and ran 100+ of them.
• Did ISO 27001 and 9001 in-house, as data protection officer, so due diligence stopped being the blocker.

2×

account size at renewal

1%

logo churn across strategic accounts

30%

shorter sales cycles

< 6 months

to ISO 27001 and 9001, done solo

Students and adult mentors message each other, so a disclosure or a grooming risk in any message had to reach a person fast.

• Moved from keyword rules to AI flags, Unitary image recognition and hash matching.
• One named person accountable for every case, on a case board with automated outreach to schools.
• Triaged every flag against KCSIE, the statutory guidance schools work to, owned the policy and reported to the board every month.

0

missed cases, every flag triaged against KCSIE

~20

cases a month, 1 in 10 at the highest severity

2.3h

average response to schools

Students joined at school, and the platform needed a reason for them to stay once they reached university.

• Growth work across the sign-up funnel and first-week activation.
• Mentee-to-mentor conversion as the bridge from school to university, so students helped at school stayed to help the next group.
• Automated with AI once the manual version worked.

6% → 40%

of mentees going on to mentor

As Zero Gravity grew from four people, hiring, budgets, revenue and support ran by hand, and each new tool added more re-keying.

• One company brain in Notion, with the tools wired together through HubSpot and Zapier.
• Every support answer written down once and every conversation tagged, so an AI assistant could take the repeat questions.
• The tags went to the product team, and later shaped the roadmap.

91%

of support conversations resolved by the AI assistant

4 → 30

people, without a dedicated support hire

2 weeks → 3 days

to onboard a new hire

£1.5m

ARR, tracked and billed through the revenue ops

About 30% of solicitor time went on fixing input errors, and the plan on the table was to hire more solicitors.

• Timed every step, and gave each of a case's 23 stages its own deadline.
• A guided intake, and integrations with the courts (HMCTS), HMRC and the banks.
• A framework for vulnerable customers, adopted company-wide.

5 → 2

solicitors needed

69%

fewer errors by case agents

2 weeks → 4 days

to handle a case

20 → 80+

NPS

Hospitals needed to feed their staff through the first COVID wave, and restaurants had closed.

• Ran it by hand on a WhatsApp group from day one, and had a working marketplace by day ten.
• A four-sided marketplace on Airtable for hospitals, restaurants, drivers and donors, with automated dispatch that halved delivery errors.
• A hiring playbook that took the volunteer team from 6 to 120.

303,000

meals in 100 days

146

hospitals, served by 223 restaurants

£1.8m

raised

97%

of deliveries on time

When I joined, operations was the founders answering emails: no CRM, no ticketing, no supplier contracts. Then the company grew 400% in a year.

• Marketing didn't go live in a market until support coverage was confirmed.
• Intercom, response templates, supplier onboarding and a crisis protocol, set up from scratch.
• ATOL compliance, safety standards and insurance, with a 100% safety record through the Sri Lanka attacks and COVID closures.

90+

NPS, held through the US launch

12 weeks

US launch plan, hit on the date

15

first ops hires, hired and trained