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.

A tutor that never hands over the answer would still earn a place in a student's week, with session limits so it never becomes a crutch.

• Coaching, practice, marking and assignments run as separate agents, each graded by its own evaluator.
• Red-teamed the failure that matters most: a student talking the tutor into handing over the answer.
• Open-coded production traces into failure modes (sycophancy on "I don't know", inconsistency across modes, hallucinated marks), then a pass/fail LLM judge on every session.
• 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. False safeguarding alarms fell 97%, with no disclosure missed.
• What it doesn't prove yet is learning. We measured use and marking accuracy, not grades. The Pioneers pilot ends with an impact report on pupils, and that's the real test.

Production came in at 4.6x the planned cost, and one model was taking 44.5% of spend. We routed by task, kept Opus on marking, and pulled a model 14x cheaper after 7 of 8 complaints traced to it.

Teachers set one homework for a whole class and only find out who was stuck when they mark it.

Roll out through teachers, class by class, rather than asking students to sign up on their own.

• Homework built from the teacher's own material.
• A weekly summary per class: who to nudge, who to stretch.
• A rollout pattern taken from the schools that worked: classes built in advance and a weekly call with the teacher.
• Data-protection paperwork ready before a school asks for it.

Help a student has to go and find mostly goes unused. So the tutor turns up inside homework their teacher has already set.

The platform held students' CVs, thousands of job posts and thousands of community posts, and no way to ask across them.

Ship retrieval and tool calling early, as a beta, so the next AI product would start with that work done.

• Tool calls across the student's CV, job posts and community posts.
• Retrieves what's relevant and talks it through in chat.
• Never invents experience the student doesn't have.

Students had nothing useful to do on the platform between mentoring sessions.

Bite-size, partner-funded courses would bring them back between sessions.

• Built with partners including Accenture, HSBC, KPMG and Snap.
• Video, bite-size progression, quizzes and an AI assignment to finish.

Students who started a pathway finished it, but they didn't come back to browse for the next one. That shaped Career Co-pilot: bring the next useful thing to the student.

A seed-stage startup selling into banks and law firms had to get through their due diligence before any programme could start.

Build each programme into the partner's early-careers hiring, and get ISO 27001 and 9001 done in-house so due diligence stopped being the blocker.

• Scoped, sold and delivered the programmes, and owned milestones and renewals.
• Took us through due diligence with HSBC, Morgan Stanley, KPMG and Mercedes-AMG F1.

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

Move from keyword rules to AI flags, image recognition and hash matching, with one named person accountable for every case.

• Regex flags first, then AI flags with Unitary image recognition, then hash matching.
• A kanban of cases, 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.

Support questions were answered by hand, and what students asked about wasn't recorded anywhere the product team could use.

Write each answer down once and tag every conversation, so an AI assistant could take the repeat questions and the product team could read the rest.

• An AI assistant answering from the team's own knowledge base.
• Bot workflows routing the rest, on SLAs of first response under 2 hours and resolution within 48.
• Conversation tags that later shaped the product roadmap.

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

Turn mentees into mentors, so students helped at school stayed on to help the next group.

• Growth work across the sign-up funnel and first-week activation.
• Mentee-to-mentor conversion as the bridge from school to university.
• Automated with AI once the manual version worked.

Hiring, budgets, revenue tracking and company knowledge had no system behind them, and each new tool added more manual work.

Build one operating system and wire the tools together, so the team did less re-keying as it grew.

• Hiring, onboarding, culture and a SaaS budget, run from one place.
• Revenue ops in HubSpot, with Zapier taking out the manual steps.
• A company brain in Notion that became the base of Zero Gravity OS.

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

Time every step, fix the input at the start and integrate with the courts, HMRC and the banks, instead of hiring.

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

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

Run it by hand on a WhatsApp group from day one, and have 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.
• NHS data-sharing agreements across the 146 hospitals, set up under crisis conditions.
• A hiring playbook that took the volunteer team from 6 to 120.

When I joined, ops was the founders answering emails: no CRM, no ticketing, no supplier contracts.

Build the operation ahead of each market: marketing didn't go live 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.