Help rebuild how commercial work in sport gets done.

You lead anzel's practice. You decide what anzel needs to know, you author the skills it sells, and you define what a good deliverable looks like.

In practice, that means you:

1. 01Author the sponsorship skills yourself, in plain language: the instructions, the worked examples, the judgement calls and the sources.
2. 02Define what “good” looks like for every deliverable, and write the reference cases each skill must pass before it ships, with the knowledge and evaluation engineer.
3. 03Decide what anzel learns next, and in what order, and keep the list of what it can do accurate.
4. 04Build a network of outside sports commercial experts and run the one-to-one sessions that turn their know-how into usable knowledge.
5. 05Lead a small team from the start: knowledge operations and a skill author, with a second practitioner to follow. Direct the AI assistants that structure expert interviews, maintain the use case list and scan the industry.
6. 06Spend time with our first customers on their first deliverables, because what a commercial director sends back is where the skills improve fastest.

You know, from running the work yourself, exactly what a director would act on unread and what they would send straight back.

In practice, that means you bring:

1. 01Eight to ten years in the commercial work of sport, and you have run it: commercial director, head of partnerships or the equivalent at a league, federation or top-division club.
2. 02Sponsorship first. You have priced a shirt or sleeve deal, written a renewal case, built a sponsor report and sent one back. Ticketing, hospitality, media or licensing beside it is a plus.
3. 03AI native: you use AI tools every day to get your own work done, you have taught others to, and you can tell an answer from a deliverable.
4. 04You can write down how you do a task clearly enough that an assistant does it your way.
5. 05You can turn forty minutes of a practitioner talking into a structured, usable entry.
6. 06Choosy about sources, and comfortable telling an engineer an output is wrong, with the reason.
7. 07No coding required. Building is: instructions, examples and test cases in plain language are the job.

You make anzel's knowledge usable by machines and turn skill quality into a measured result, not a discussion. Nothing ships until it clears the bar.

In practice, that means you:

1. 01Build the evaluation harness, hands on. Until our senior agentic engineer joins you stand it up yourself: test runs on every change, results anyone can read, and a release that stops when the bar is missed.
2. 02Build the reference test sets and grading rubrics every skill must pass, with the practitioner, so the bar for “good” comes from evidence, not opinion.
3. 03Turn the corpus into something machines can use: chunking, indexing, metadata and versioning that lets every change be traced and rolled back, with retrieval quality measured, not assumed.
4. 04Write every failure reported from the field into the test set before the fix ships, so the same mistake never ships twice.
5. 05Track pass rates over time and publish the standing comparison between a general model and anzel on the same commercial question, both graded, with the gap as a number that improves every quarter, and what each piece of work costs beside it.
6. 06Direct the AI assistants that do the recurring work: the graders that score every output, the regression monitor that reruns every test when anything changes, and the checks on extraction quality.

You can build automated evaluations people trust, because you know exactly where they disagree with human judgement, and why.

In practice, that means you bring:

1. 01Experience building evaluation sets and graders for language model systems in production, calibrated against human judgement.
2. 02A feel for where model-graded evaluation lets bad work through: position and length bias, graders that prefer their own family's writing, rubrics that reward form over substance.
3. 03Experience building retrieval pipelines, and the conviction that retrieval quality, not model choice, is usually where output quality is won or lost.
4. 04You write the code yourself and work with coding assistants as a matter of course. You can stand up a harness alone and hand it over cleanly later.
5. 05Careful data modelling instincts: taxonomies, metadata, versions. Careful with small samples, and clear about how sure you are.
6. 06Comfortable spending an afternoon with a domain expert and coming out with a rubric they recognise as their own. You would rather find a failure yourself than hear about it from a customer.