A Blog by Robin Howlett
I'm Robin Howlett, Software Development Manager at Amazon Advertising. Married to Sarah Howlett. Father of twins. Boulder, Colorado-based. Made in Ireland.
The 2026 Breeders’ Cup is at Keeneland on October 30 and 31, and I’ve launched a game for it. 14–0 is free: you draft one horse for each of the 14 Breeders’ Cup races from every running since 1984, then sim the card, watch each race, or ride it yourself, at the track where that year’s race was run.
In this post, I explain where it came from, how it works, and how it was built.
In June 2026, the thoroughbred world was sent spinning when confidential veterinary health data leaked onto X (formerly Twitter), displayed within augmented past-performances posted by a horse racing social media rabble rouser. It took weeks of frantic whispers, preliminary denials, and a digital manhunt before the story finally broke wide open: the source of the leak wasn’t an organized crime syndicate or a disgruntled groom, but Dr. Marshall Gramm, a Rhodes College economics professor, founding partner of Ten Strike Racing, and a ubiquitous figure across handicapping contests, claiming ranks, and industry governance committees. Gramm wasn’t just a casual observer; he was a horse owner, bettor, and had a media profile, an “insider” who had spent years analyzing racing information and participating in the very types of high-level industry circles where data privacy and oversight were hammered out.
For weeks after those custom past-performances for Deterministic and Griffin’s Wharf surfaced on X, posted by Justin Wunderler, who then started pulling the thread, HISA’s public line was that nothing had gone wrong on its end. On June 15, CEO Lisa Lazarus said flatly that the data “could not have come from the HISA portal.” It had. Reconstructing what happened took roughly two months of digital forensics, contractor interviews, and a third-party cyber investigation by the firm Arete. On August 17, 2026, HISA filed two disciplinary charges against Gramm, one under the rule governing access to veterinary records and one for fraud.
What tips this from a data-policy case into something more surreal is a single scene. On the morning of July 17, Gramm joined an hour-long call convened by The Jockey Club, to which he had been elected in 2024. The subject was the leaked past-performances and whether the industry should start disclosing veterinary records to the public. On the call with him were Lazarus, Patrick Cummings of Mike Repole’s National Thoroughbred Alliance, and Jockey Club executives Jim Gagliano and Charlotte Clément. Gramm offered ideas. He did not mention that he was the reason the meeting was happening. Around 5 p.m. that same day, HISA investigators confronted him, and he acknowledged being the source. Lazarus already knew where the trail led while she sat on the morning call. Nobody else on the call did. A man was helping the industry craft its response to his own leak.
I introduced redboarder, a place to practice the craft of handicapping with an AI partner, in a separate post. This one goes under the hood: not what it does, but why each piece is shaped the way it is.
Most of the interesting decisions in a product are invisible in the finished thing. You only see them if someone tells you what the alternatives were and why they lost. So this is a walk through the main ones, roughly in the order you’d meet them using the app: why the races are old, which ones qualify, why I built my own rating system, why the AI sits where it does, how you bet, how the results stay honest, and why the AI operates under such tight rules.
Nearly all of it comes back to two masters that every decision has to serve at once: keep it honest, so the practice is real, and keep it cheap enough that one person can run it.
Becoming a successful horseplayer is extremely difficult. Interpreting form, pace, class, and the myriad of other variables that can affect the performance of a living creature, and turning that read into a wagering decision that reflects the value your opinion identified across a host of different bet types - well, it’s one of the last forms of betting where a sharp individual can still find an edge against both the “house” (takeout etc) and the crowd and profit handsomely when they are right.
Improving these skills, however, is more challenging than it should be, for a handful of reasons that pile on top of each other.
Start with the structure. Pari-mutuel wagering, the way it works in America, is zero-sum by definition: everyone bets into the same pool, so your edge comes straight out of someone else’s payout. Which means nobody who’s actually good at this has the slightest incentive to help you get better. The loudest voices in the space are tipsters, and their business model is keeping you dependent on their picks instead of building your own judgement.
Then there’s where the attention goes. Winning at the races is, in the end, a betting problem. The edge is thin, and what little of it exists lives in the dry, objective side of the craft: value, pool dynamics, staking, discipline. That side is unglamorous and binary (you either had the value or you didn’t), so it gets almost none of the airtime. What gets debated endlessly, on every handicapping show and in every tip sheet, is the opinion: who the best horse is. It’s subjective, it’s arguable, it’s fun, and it’s the part that matters least once you can already read a race. Newcomers pour their hours into the entertaining half and starve the half that decides whether they win.
And the cost of learning the hard way is brutal. Pari-mutuel takeout, the house’s cut of every pool, dwarfs the vig at a sportsbook, so every bet you place while you’re still bad is fighting a steep built-in edge before you’ve even been wrong. Put the takeout, the tipster noise, and a genuinely thin edge together, and a beginner betting live money mostly just watches the bankroll bleed away as tuition. Even if you’re happy to pay that tuition, you can’t practice efficiently. Live racing trickles out one race every thirty or forty minutes, most of them cards you don’t care about, with no way to drill when you actually have a free hour.
So the problem was never a lack of information. Old races have known outcomes, so replaying them teaches you nothing your hindsight doesn’t already know. Tipsters just hand you the pick, so following along tests nothing. You can read forever and still have no honest signal on whether you’re any good.
I spent years building the raw material for this without quite admitting it was the gap I kept circling: parsing Equibase chart data, building Handycapper, and assembling a database of more than a million races. The industry has talked about technology for as long as I’ve been around it, and not much of it ever shipped. It turned out an independent with the data and a few ideas could just build. But data alone doesn’t sharpen you; it’s something you query, not something that makes you better. What I wanted was a place to make a real call and find out, honestly, whether I was right, and along the way to see what a bit of genuine product thinking could do in a corner of the world that hasn’t seen much of it.
So I built it. It’s called redboarder, and this post is a look at what it is: who it’s for, how it works, what’s in it, how it was built, and why I think it’s interesting.
When I added an AI partner to redboarder, a place to practice the craft of handicapping, the obvious move was the one nearly everyone reaches for when they want to “chat with their data”: point the model at the database, or embed everything and let retrieval feed it context, and let it answer. I did the opposite, and I’ve come to think the opposite is right far more often than the default suggests.
The AI in redboarder never touches the data directly. It can’t query the database, it can’t do the arithmetic, and it never sees a result it isn’t supposed to. It’s handed a finished, structured brief and asked only to reason and talk about it.
The model is the narrator, not the analyst.
In this post I’ll make the case for that split: why it’s the right way to put this kind of data behind a model, and how it’s the single decision that makes the AI convincing, cheap, correct, and impossible to trick into spoiling the game all at once. And, at the end, where it would be the wrong choice.