Human-in-the-Loop Agentic AI Workflow Design Product Architecture

Clear
Signal

An agentic AI pipeline that ingests a dense horse racing program, generates a clean expert pick sheet, translates it into three languages, and delivers it timed to each market’s morning window, with one human approval step before anything goes out.

3
Markets served
3
Languages
1
Human step
<2m
To a decision
Problem
01 — Context

The data is
the barrier

A Daily Racing Form 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.

The document
Dense, expert-coded, built for insiders
Daily Racing Form PDFs assume literacy in speed figures, class ratings, trainer codes, and form cycles. Years of convention with no on-ramp for casual or international bettors.
The audience
Shifted online, increasingly international
US casual bettors, Latin American fans with growing digital access, European markets — all willing to pay, none able to parse the source document.
The expert
Can read it. Can’t scale distribution.
The expert has the picks and the voice. What he doesn’t have is infrastructure to reach an international audience with a polished, timely product every racing day.
The gap
Nothing between the form and the bet
No product exists in this space: a clean, expert-authored, translated, timed pick sheet positioned as a subscription with a racetrack distribution partner.
Reframe
02 — Design Framing

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.

Before
What most tools do
Reproduce the raw data in a different format
Assume the user can interpret what they’re shown
Deliver in one language, one timezone
Require the expert to build the output manually
Fully automate or fully manual — no middle path
Clear Signal
What this system does
Parses and ranks data so only the signal reaches the subscriber
Delivers a decision, not a dataset
Three languages, three timezone-targeted sends per day
Expert approves; the system builds, translates, and sends
One deliberate human gate before any distribution
System Design
03 — Agentic Pipeline

Seven steps.
One deliberate pause.

Designed and built in n8n, 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.

01
PDF Ingest
The Daily Racing Form PDF triggers the workflow. Structured extraction pulls race-level data — horse, jockey, trainer, class, speed figures, form cycle — into clean objects the downstream agents can reason over.
n8n triggerStructured extraction
02
Parse & Rank
An AI agent applies our expert’s ranking logic: speed figures, class drops, connections, pace scenarios, track bias. The output is ranked picks per race with supporting rationale, not raw data, but processed judgment.
AI agentExpert logic encoded
03
Generate Layout
Picks, confidence tiers, and plain-language rationale are rendered into the Clear Signal template. The layout is designed to communicate a decision, not to educate. A subscriber should be informed in under two minutes.
Template engineConfidence tiering
04
Expert Review & Release
The expert receives a formatted preview. The expert can approve, flag a change, or hold. Nothing sends without his release. This step costs him almost nothing on a clean day — and catches everything that would cost him credibility if it went out wrong.
Human-in-the-loop: by design, not default
Human gate ✦
05
Translation
The approved sheet is translated into Spanish and Portuguese with racing-specific terminology preserved. General translation breaks on domain vocabulary — the system is tuned to understand the difference between “class” as a concept and “class” as a racing term.
Spanish · Portuguese · EnglishDomain-aware NLP
06
Timezone Scheduling
Three sends staged to land in each market’s morning window before racing begins. A pick sheet that arrives after the first race is worthless. The value is entirely time-sensitive. This step isn’t logistics; it’s the product working correctly.
ET · BRT · CETMarket-timed delivery
07
Distribution
Email delivery to subscribers, segmented by language and market. Co-branded with the racetrack partner. Revenue split between the expert, the track, and the platform. The track has an incentive to promote; our expert has an incentive to maintain quality; subscribers have a reason to stay.
Email deliveryCo-brandedSegmented
Key Decisions
04 — Design Rationale

Why each choice
was deliberate

Decision 01
Human-in-loop over full automation
Full automation was technically feasible. It wasn’t right. The expert’s value — what subscribers pay for — is his judgment, not just his data. A fully automated sheet is a product without an author. The one-click approval preserves his voice, his credibility, and his ability to catch a scratch before it becomes an embarrassment. Racing is live. The gate isn’t a bottleneck; it’s a feature.
Decision 02
One pipeline, three languages
The international opportunity is a distribution problem disguised as a language problem. The picks are the same. The analysis is the same. What changes is language and delivery time. Building separate products per market would fracture the workflow and multiply the expert’s workload. One pipeline, three outputs. The insight scales; the labor doesn’t.
Decision 03
Layout as decision, not education
The Daily Racing Form educates. Clear Signal orients. There’s a difference. The layout strips everything that isn’t a decision: race number, pick, plain-language rationale, confidence tier. Three tiers (HIGH, MED, LEAN) let subscribers calibrate without needing to understand the underlying analysis. The expert’s judgment surfaces as signal, not data.
Output Design
05 — The Sheet

Confidence in
under two minutes

The Clear Signal 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.

Clear Signal
Keeneland · Apr 5, 2025
Expert Selections
🇺🇸 EN  ·  🇲🇽 ES  ·  🇧🇷 PT
1
#4 Midnight Logic
Class drop off a bullet work · wet bias favors front-runners today
6F / Dirt
High
2
#7 Coastal Wind
Best Beyer in the field · proven on this turf course
1M / Turf
Med
3
#2 Iron Veil
Wide draw hurts · pace sets up perfectly if she breaks clean
1⅛M / Dirt
Lean
4
#1 Sola Fide
First off claim · undervalued at current odds — watch the tote board
5½F / Turf
High
© Clear Signal · Keeneland Partnership Delivered 8:02 ET · 9:02 BRT · 14:02 CET
Clear Signal · Output Preview · English Edition
Stakeholders
06 — Business Model

Three parties.
Aligned incentives.

The system works because no one is being asked to do something that doesn’t serve their interests. The design holds the business logic together the same way it holds the workflow together.

our expert
Expert · Voice · Approver
The expert’s picks and credibility are the product. The pipeline amplifies his reach without multiplying his effort. Our expert reviews; the system builds, translates, schedules, and sends. The one-click approval is also his quality gate — the moment his judgment enters the product before it reaches a subscriber.
The Racetrack
Co-brand · Distribution Partner
Co-branding ties Clear Signal to the venue’s credibility. A revenue share gives the track a direct incentive to promote distribution to their existing audience. They provide legitimacy and reach; the system provides a product worth promoting.
Subscribers
US · Latin America · Europe
Casual and international bettors who want a confident, readable recommendation in their language, in their morning window, before race day begins. They’re not paying for data. They’re paying to skip the part they can’t do themselves.
My Role
07 — Contribution

From domain insight
to working system

↗
Product Architecture
Conceived the pipeline structure: what the system needed to ingest, what it needed to produce, where human judgment needed to stay in the loop, and how the business model closes around the workflow.
↗
Agentic Workflow Design (n8n)
Designed and built the full n8n workflow: PDF ingest, AI parsing and ranking, layout generation, human-in-the-loop approval, translation, timezone scheduling, and segmented email distribution.
↗
International Distribution Strategy
Identified the Latin American and European audience opportunity as a structural distribution gap, not a localization problem. Designed the timezone delivery logic that makes the product viable across markets without additional expert labor.
↗
Human-in-the-Loop Approval System
Designed the review-and-release interface so the expert’s approval step is frictionless. The friction is calibrated: low enough that he approves without resistance; present enough that he actually reviews. One action. Everything else is the system’s job.
↗
Clear Signal Layout Design
Designed the output format around one constraint: decision in under two minutes. Three confidence tiers. Plain-language rationale. No jargon, no codes, no glossary required. The expert’s judgment surfaces as signal, not as data the subscriber has to interpret themselves.
Outcome & Opportunity
08 — So What

A product that
didn’t exist.

Clear Signal started as a distribution problem and became a product design problem. The Daily Racing Form was never going to change. The international audience wasn’t going to learn to read it. The gap was real, and the infrastructure to close it — agentic AI, multi-language NLP, automated scheduling — was newly available.

The design work was in knowing where not to automate. The one-click approval step is the most important element in the system — the point where the pipeline’s output and the expert’s judgment meet before anything reaches a subscriber.

The deeper opportunity is scale. The pipeline architecture doesn’t care which racetrack it’s producing for. A system designed for Keeneland is a system that can run for any track with a racing form and an expert willing to put their name on the picks. Clear Signal is a product. It’s also a platform.

3
Markets served

US, Latin America, Europe — one pipeline, three languages, three timed delivery windows.

1
Human step

Expert approval is the single non-automated step, by design, not by default.

<2m
To a decision

Subscribers go from inbox to informed pick in under two minutes, in their language.

“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.”

Design principle · Clear Signal