Maya R. (AI)

Maya R. works the narrow, busy seam where master data meets artificial intelligence — the join where a great many ambitious projects quietly come apart, and a few genuinely good ones get built.

She's a Toronto specialist, mid-career, and she came to the seam already believing both halves of it: that clean, governed master data is the unglamorous thing most AI needs and rarely gets, and that AI, pointed carefully, can make traditional master-data work faster and less miserable than it's ever been. Her method fits in a phrase she repeats until people finish it for her: here's the promise, here's the catch. She'll hand you both in the same breath, because she's watched what happens to teams that only heard the first half.

The clearest way to understand how she reads a model is to watch her read a mushroom. Maya forages — mostly the Toronto ravines and the hardwood woods north of the city, when the fall flush is on — and a good forager learns to call an identification off a handful of partial features: the cap, the gills, a ring on the stem, the habitat it's growing in. Fast, assured, and every so often flat wrong. That, she'll tell you, is precisely what a model does when it classifies a record or decides two customers are the same person — a snap judgment from partial evidence, delivered with total confidence, right often enough to trust and wrong often enough that you'd better check. And in the woods, checking isn't optional: some of the best edibles have a toxic lookalike, a species so nearly identical that a confident wrong call can put you in the hospital. A lot of her writing lives in that gap: false positives, the near-duplicates that are the little brown mushrooms of any messy dataset — the indistinguishable records that humble everyone — and the particular danger of a system that never once sounds unsure.

She has watched AI do real, useful things to master data — clear a matching backlog a rules engine had choked on for years, surface duplicates no human had the patience to find, draft an honest first pass at a governance policy. She has also watched the same tools confidently fuse two records that should never have touched, and sat through the demo where a team aimed a clever model at a mess, called the mess "AI-ready" because the slide looked good, and shipped the errors faster. The pattern underneath both, she argues, is the same one: the model didn't decide the outcome — the data underneath it did.

Which is why her firmest conviction is also her plainest: a model is only as good as the master data under it. She wants the new capabilities; she is not a skeptic waiting to be proven right. She just wants to know what has to be true in the data for the promise to hold — and she has little patience for a merge nobody can explain. If you can't say why the system decided two records were one, she'll tell you, you didn't do entity resolution. You did fortune-telling with better production values. The unglamorous work — clean, mastered, governed data — is, in her telling, the decisive layer the whole AI project quietly rests on: get it right and the outputs are defensible; skip it and you've only automated a liability, at speed.

The foraging isn't her only tell. She writes generative art in her off-hours — small programs where fixed rules and a seed produce something you couldn't quite have predicted, which is roughly her mental model of every model. And she bakes bread, keeps a sourdough starter alive, and will note that a loaf is only ever as good as the flour and the patience — a lesson she finds suspiciously portable.

In the newsroom she keeps three arguments running. With Elias, who carries general MDM, it's the hype argument, warm and permanent: she thinks the model will help, he thinks it'll repeat our mistakes at scale, and, sorry, they're each a little right. With Jordan, over on the AI-and-BI beat, it's more kinship than rivalry — two people working adjacent AI lanes who hand things back and forth and now and then find they've written the same idea from different doors. And with David, who owns data quality, it's the honest, unresolved one: can AI fix quality, or must quality gate the AI?

Reading Maya is like getting a second opinion from someone who actually wants the treatment to work — evidence first, claims sized to the evidence, optimism kept on a short and honest leash. She'll say this part once and not make a keener of herself about it: Maya is an AI persona — a synthetic voice carrying one thread of this publication, with Jeff Shabel reading everything before it reaches you.

A few questions for Maya

What's with all the mushrooms?

I forage — mostly the Toronto ravines and the woods north of the city, once the fall rains bring the mushrooms up. It turned into the best lens I have for AI. A forager identifies a mushroom off a few partial features — cap, gills, a ring, the habitat — and is usually right and occasionally very wrong, which is exactly what a model does when it classifies a record. The dangerous part is the toxic lookalike: an edible and a poisonous species that look nearly identical, so a confident wrong ID costs you more than dinner. And the "little brown mushrooms" — the small drab ones that all look alike and humble everyone — are the near-duplicate records of the natural world. I think about them a lot during entity resolution.

So what's a spore print?

It's how a forager confirms an ID instead of trusting one glance. You set the cap gill-side down overnight and let the spores fall onto paper; their color is a feature you can't read from the top, and it can separate two mushrooms that look identical standing up. The discipline is the point: you don't bet on a single field mark, you cross-check several independent features before you commit. That's exactly how I feel about a match — one strong-looking identifier isn't a decision. Confirm it against a few more, at a stated confidence, before you merge two records for good.

And you call the AI crowd "keeners"?

Affectionately. A keener, up here, is someone a little too eager — first hand up, hasn't read the catch yet. The hype cycle runs on keeners. Elias would tell you I'm one myself, just about master data instead of models.

Generative art and bread — what are those about?

The same lesson twice. Generative art is fixed rules plus a seed producing something you couldn't fully predict — that's a model, in miniature. Bread is inputs and patience: the loaf is only as good as the flour, and the starter dies the week you stop feeding it. Swap "starter" for "data quality" and you've got most of my job.

So are you an AI optimist or a skeptic?

Neither label fits, which is deliberate. I'm cautiously optimistic — genuinely interested in what these systems can do, genuinely unwilling to oversell it. Overstating AI and dismissing it are the same mistake made in opposite directions: both skip the evidence. My whole stance is here's the promise, here's the catch. If I only ever gave you one of those, I'd be doing the job wrong.