Job Matches started as a real job-search workflow. We turned it into a live UK product.

**Job Matches** is AI Build Group's live UK job-matching product. We created it from a real job-search workflow, then deployed it as a multi-user platform at [job-matches.co.uk](https://www.job-matches.co.uk/): candidates upload a CV, set criteria, review matched roles, and receive a tailored CV and covering letter for each job they choose to pursue.

This is a **Job Matches case study** from our consulting practice, not a product brochure. Claude helped us take a named workflow from personal MVP to production software, with clear ownership, data boundaries, and human review on applications. AI Build Group owns and operates the product as part of the Job Matches document and careers stack.

What is Job Matches?

Job Matches is a UK job search platform. It matches a candidate's CV and preferences to roles, then prepares application documents for the jobs they select. Candidates start free; paid plans unlock more matches. Employers can list a role or search an opted-in CV pool, priced below comparable UK job boards on the public product site.

• **Candidates:** upload a CV, set roles, location and working pattern, then review matches. Each match can include a tailored CV and covering letter.
• **Applications:** the candidate applies. Job Matches does not submit applications automatically.
• **Employers:** post a role for 28 days with daily CV matches, search opted-in CVs, or combine both on a hire plan.
• **Privacy default:** employer discovery is off until the candidate turns on Allow employers to find me.

Those rules matter for consulting buyers. An AI product that applies for jobs without a person in the loop would be faster to demo and harder to defend. We kept the high-consequence action with the user.

Why did we build a product instead of another slide deck?

The first version was a personal finance job-search engine: collect listings from compliant sources, filter and deduplicate before any model call, classify relevance, group roles into reusable CV datasets, and generate one CV variant per mature dataset. That workflow ran on demand and on a daily London-time schedule.

It was useful because it removed a recurring dependence: the same search, ranking and document preparation every day, without removing human judgement on which roles to pursue. That is the same weekly habit described in I Make Myself Redundant. Consulting value starts when that habit becomes a system other people can use.

Version 2 is the commercial platform: multi-user accounts, a public website, search against live UK roles, career profiles, application packs, employer listing, billing, and production hosting. We did not rip out the original engine; we isolated it and rebuilt the product around the same job.

How did Claude help us create and deploy Job Matches?

Claude was the primary implementation partner for the production build. AI Build Group is a Claude Partner; this story is how we used that partnership on our own product, with the same delivery standard we use for client AI assistant and workflow re-engineering work.

The work was not a single prompt. It was a governed agent loop: named jobs, repository context, tests, cloud configuration, and human review of product decisions. Typical jobs included:

• Turning the personal search workflow into a multi-user product without sharing the original database or secrets.
• Designing candidate and employer journeys so applications stay human-reviewed.
• Wiring CV ingest and Word application packs through OfficeMaker rather than ad-hoc document scripts.
• Deploying the web app, identity, data store and search worker on AWS, with secrets in SSM rather than in source.
• Keeping ranking cheap where rules suffice, then calling a model only after filters and deduplication.

That last point is a consulting control, not an optimisation footnote. Model calls after deterministic filters reduce cost, reduce hallucinated matches, and give IT a story they can defend. The same pattern appears when we pair assistants with ChatGPT Business workspaces: the model is not the whole system.

What architecture did we ship?

Job Matches is a production stack, not a notebook demo. The public product runs as a Next.js application on AWS Amplify Hosting, with Amazon Cognito for sign-up and sign-in, DynamoDB for profiles and matches, and an App Runner search worker for live role discovery. Daily email digests use Amazon SES.

Career profiles store a structured extract of the uploaded CV. Matching ranks live roles against that profile and stated preferences. Where OfficeMaker is configured, Word uploads convert to structured document JSON; per-job CV and cover-letter packs are created as Word files on governed templates. Conversational letter drafting can sit alongside those packs through Letters4u in the same AI Build portfolio.

This is the two-layer pattern in Embedded AI and Central AI Assistants: intelligence inside the product people use, plus a central document service that other systems can call. Job Matches is the embedded product. OfficeMaker is the document capability. Claude helped us design both layers so they stay separable.

How do matching and documents stay governed?

Governance here is product behaviour, not a policy PDF.

• **Source discipline.** Discovery prefers official APIs, feeds and public careers routes. The platform does not bypass login walls or submit applications.
• **Filter before generation.** Ranking and document tailoring run after profile and preference filters, not instead of them.
• **Human application.** The candidate reviews matches and applies. That is the review gate for employment actions.
• **Employer access is opt-in.** CV search only covers people who allowed employers to find them.
• **Document service account.** Application packs go through OfficeMaker on an AI Build tenant, so Word generation follows the same schema-first path we use in client document programmes.
• **Secrets off the repo.** Production keys live in AWS SSM SecureString parameters, not in application source.

UK GDPR still applies: Job Matches processes candidate CVs and contact details. The operating rule is the same one we use in AI governance consulting: know what the system may read, write and send, who owns it, and where a person must still decide.

What did deployment actually include?

Deployment meant the product is usable by people who are not the original job-seeker. That required identity, billing, employer flows, email, admin access, and a public domain, not only a ranking model.

• Public site and authentication at job-matches.co.uk.
• Candidate dashboard, CV upload, search preferences and match review.
• Optional application document packs via OfficeMaker.
• Employer registration, role posting and opted-in CV search.
• Stripe billing for candidate plans and employer products.
• Operational paths for daily digests and internal admin status.

Public employer prices at launch (as shown on the product site) are £60 for a 28-day role with daily CV matches, £111 per month for CV Search, and £100 per month for Hire, which combines a live role, daily matches and CV Search. We publish those figures because they are on the live site; we do not invent user counts or time-saved percentages for this story.

What should UK buyers copy from this build?

If you are buying AI consulting because you want a product, not a workshop, this is the sequence we would reuse.

1. Start from a workflow someone already feels

Job search was slow, repetitive and document-heavy. We did not begin with a platform strategy. We began with one person's daily queue.

2. Separate the MVP from the product

The personal engine stayed isolated. The commercial platform got its own database, identity, and secrets. That is how you avoid turning a useful prototype into an unsupportable production risk.

3. Put Claude on named implementation jobs

Use Claude to design, write, test and deploy against a repository and a definition of done. Do not treat the model as an unbounded chatbot sitting on customer data.

4. Keep humans on high-consequence actions

Job Matches prepares matches and documents. People apply. The same rule belongs in HR, finance and customer-reply assistants.

5. Generate documents through a governed service

OfficeMaker gives Job Matches a schema, templates and an audit trail for Word packs. That is cheaper to operate than one-off Python document scripts, and it is the same capability we deploy for other UK teams.

How does this support AI assistant consulting?

Buyers asking for an AI assistant for business often want proof that we ship systems, not only prompts. Job Matches is that proof inside our own portfolio: a named job (match CV to UK roles), an owner (AI Build Group), data boundaries (opt-in employer access), a review gate (the candidate applies), and a production URL.

It also shows when not to force everything through a central chat tab. Matching and document generation belong inside the product. A central assistant can still help with career questions or letter drafting through Letters4u. That is the architecture we recommend: intelligent applications below, orchestration where work crosses systems.

Sources and related pages

• [Job Matches](https://www.job-matches.co.uk/) — live product
• Job Matches on AI Build — product owner page
• Consulting — how we scope delivery
• AI assistants — four-week assistant rollouts
• OfficeMaker — governed document generation
• Results and case studies — measured client programmes

If you want the same method on a workflow your team already owns, book a discovery call. If you want to try the product we shipped, start on [Job Matches](https://www.job-matches.co.uk/).