From domain insight
to working system
My client came to me wanting a one sheet for bettors. When we mapped out every step from raw racing data to finished picks, it became clear this was a distribution system, not a design job. The pipeline pulls structured data from a dense racing PDF, formats it into clean, plain-language picks, and stages it for delivery. The single human step exists because race conditions change on the day, and nothing goes out without approval. His credibility is on the line every race day. That gate isn't a workaround. It's the product.
The data is
the barrier
A raw racing data PDF is a masterpiece of compression. Decades of convention packed into columns of abbreviations, numbers, and codes that mean something specific to someone who spent years learning to read them. For everyone else, it’s noise.
The horse racing audience has migrated online and gone international. A significant following in Latin America and Europe wants to participate, but the information infrastructure still assumes an American expert in the grandstand. The problem isn’t that the analysis is hard. It’s that the starting point requires fluency most bettors don’t have and can’t easily acquire.
Not a data problem.
A translation problem.
The racing form has the right information. The pipeline doesn’t need to create insight. It needs to surface it. That reframe changes everything about how the system is designed.
Seven steps.
One deliberate pause.
Designed and built in Python, the pipeline is fully agentic from ingest through staging. The single non-automated step isn’t an oversight; it’s the most important design decision in the system.
Why each choice
was deliberate
“Good agentic design isn’t about automating everything. It’s about knowing exactly which step deserves a human, and making that human’s job as effortless as possible.”
Confidence in
under two minutes
The Track Intel sheet is designed around one constraint: a subscriber should be able to open it, scan it, and feel oriented before coffee goes cold. Every element either earns its place or gets cut.
A product that
didn’t exist.
The pipeline architecture doesn’t care which racetrack it’s producing for. Any track with a racing form and an expert willing to put their name on the picks can run this system. Track Intel is a product. It’s also a platform.
The design work was in knowing where not to automate. That’s the whole job.