The AI that runs Startline
Run by AI and automation.Every job, in the open.
Startline is built by one founder, Ricky Ybarra. The day-to-day work is done by software: it re-reads rosters, checks coach lists, finds camps, reads school news and sends what families signed up for. This page lists every one of those jobs: what it does, how often, when it last finished, and which ones use AI.
- 23
- jobs run by themselves
- 6
- of them use AI
- 17
- are automation, with no AI
Dates read from the job record · the watchdog last ran on Oct 6, 2026
The scores are counted, not generated.
The parts a family decides on are arithmetic. No AI writes a score, a roster count or a price, so the same inputs give the same answer every time.
The Fit Scores
Arithmetic on a player's answers and each school's data: the same inputs always give the same two scores, and every point is explained on screen.
The roster counts
Counted from rows read off each program's own roster page. Nothing is estimated or filled in.
The first email draft
Built by fixed rules from the family's answers and one true, sourced fact about the school. A model only changes the wording, and only if the player recorded a voice.
Reply help and follow-ups
A coach's reply is sorted by a table of patterns, and the follow-up plan is a set of dated rules.
What it adds up to,for one player.
One sample player, 103 seconds. The questions, the schools that fit and why, the roster at their position, the first email, the plan.
Open the film on its own pageKeeping the data current
Rosters, coaches, camps, news and results, re-read from the schools' own sites.
- Automation · no AI
Re-reads every college roster
Opens each program's own roster page and reads every player's position and class year with fixed parsers. No AI is used on this run, so a count is a count.
- How often
- The 1st of every month, and every Monday in April, from mid-August through September, and in December
- Last finished run
- Oct 1, 2026
- What is on file
- 66,998 roster spots on file across 2,215 programs. 2,194 programs are confirmed on their 2026-27 squad. Counted Oct 6, 2026.
- When it is unsure, or fails
- A program's rows are replaced only if the validator clears them and the whole run passes its gates; otherwise the old rows stay. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Reads the rosters that block cloud servers
Some college sites refuse requests from cloud servers. Their rosters are re-read from one Mac instead, with the same fixed parsers.
- How often
- Sundays, early morning Pacific time
- Last finished run
- It leaves no record this page can read, so no date is shown. It runs on the schedule shown.
- When it is unsure, or fails
- It writes roster rows and a record of which season each squad is, and nothing else. It stands down if the main roster run is owed or still running, so the two never write at once.
- Automation · no AI
Publishes what the Mac read
Checks once a day whether the Mac wrote new rosters, and if it did, rebuilds the site's season record from the database so those school pages say which season they show.
- How often
- Every day
- Last finished run
- It leaves no record this page can read, so no date is shown. It runs on the schedule shown.
- When it is unsure, or fails
- It reads no page and writes no roster row. The rebuilt files go live only if the roster run's own gates pass on them.
- Automation · no AI
Checks every coaching staff list
Re-reads each program's staff page and compares it with the coaches on file: who left, who arrived, whose address changed.
- How often
- Thursdays from January to June, and the 8th of the month from July to December
- Last finished run
- Oct 3, 2026
- What is on file
- 7,548 coaches on file, 5,517 of them with an email address. Counted Oct 2, 2026.
- When it is unsure, or fails
- Anything that looks wrong is held and changes nothing. Every program's previous rows are kept so a change can be undone, and a run that changes too much at once changes nothing at all. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Re-checks every ID camp
Re-reads each listed camp's registration page and checks its dates. It adds next season's edition of a camp once the school's page has printed the new date twice, at least five days apart.
- How often
- Wednesdays from January to July, and the 15th of the month from August to December
- Last finished run
- Sep 25, 2026
- What is on file
- 1,888 ID camps listed at 1,208 schools; 812 have a date posted.
- When it is unsure, or fails
- Dead links, cancellations, moved dates and changed prices are held, not published. It never edits or removes a listed camp. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Finds new ID camps
Walks each program's own athletics site for camp pages and reads the camps out of their structure: a soccer ID camp, run by the school, with a future date and a registration link that answers.
- How often
- Every Tuesday
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- A camp is added only on a second sighting at least five days after the first. Everything it read but would not add goes to a review file, never to the directory. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Collects soccer headlines
Reads each program's own athletics news feed and each university's newsroom feed, and keeps a headline, a date and a link for each story. Never a story's text or pictures.
- How often
- Mondays and Thursdays
- Last finished run
- Oct 6, 2026
- What is on file
- 8,370 headlines on file, from the feeds of 1,899 programs.
- When it is unsure, or fails
- The new file goes live only if the type check and the whole test suite pass on it. Nothing is written to the database.
- AI
Reads school news
For the schools families follow, an AI model reads the school's own news page and writes one sentence a recruit could mention, with the link it came from. If it finds nothing it can point to, it writes nothing.
- How often
- Every day
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- Only a sentence the page states plainly can open an email. Anything the model was less sure of is held and is not used. If a due run is missing or fails, the daily watchdog emails the founder.
- Model
- Claude Sonnet 5, by Anthropic
- AIas a backup
Reads match results
Reads each followed team's latest result. A fixed parser reads the page first; an AI model reads it only when the page's layout defeats the parser. In the autumn a second run does the same for youth clubs.
- How often
- Every day; the youth-club run from August to November
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- A result with no completed match behind it is not written. If a due run is missing or fails, the daily watchdog emails the founder.
- Model
- Claude Opus 4.8, by Anthropic
- Automation · no AI
Collects the week's college scores
Opens the NCAA's scoreboard in a real browser and stores the week's fixtures and results, so the weekly update has this week's games in it.
- How often
- Every Monday
- Last finished run
- Oct 5, 2026
- When it is unsure, or fails
- A game already stored is updated in place, never duplicated. If a due run is missing or fails, the daily watchdog emails the founder.
Working with a family's own words
These run only when a family asks for them.
- AI
Learns how a player talks
A player talks for about two minutes. An AI model reads the transcript once and writes a voice profile: tone, energy, and the player's own phrases, quoted word for word.
- How often
- When a player records a voice intro
- Runs
- Only when someone asks. There is no schedule.
- When it is unsure, or fails
- The family sees the profile on screen and can rebuild it. If the model is unavailable, a basic profile is built from the player's own words by fixed rules, and the screen says which one it is.
- Model
- Claude Opus 5.5, by Anthropic
- AI
Rewrites a draft in the player's voice
The email draft itself is built by fixed rules from the family's answers and the school's roster. If the player has a voice profile, an AI model rewrites the wording to sound like them. The facts, numbers and links do not change.
- How often
- When a signed-in family drafts an email
- Runs
- Only when someone asks. There is no schedule.
- When it is unsure, or fails
- A rewrite that loses the film link is thrown away and the standard draft is shown. The family reads and edits every email and sends it from their own inbox. Startline never sends one.
- Model
- Claude Opus 5.5, by Anthropic
- AI
Reads an uploaded schedule
A family uploads a PDF or a photo of their schedule. An AI model reads the dates, opponents and places out of it.
- How often
- When a family uploads a schedule
- Runs
- Only when someone asks. There is no schedule.
- When it is unsure, or fails
- The dates it read are listed for the family to check, and what the family typed always still counts.
- Model
- Claude Sonnet 5, by Anthropic
- AI
Answers recruiting questions
Ask Startline is the assistant in the corner of the site. An AI model answers from Startline's counted data and guides, looks a school up before it says anything about it, and links the page each answer came from.
- How often
- When a visitor asks
- Runs
- Only when someone asks. There is no schedule.
- When it is unsure, or fails
- It cannot see or give out a coach's contact details, it keeps no conversations, and daily caps pause it before it can run up a bill. No person reads its answers, so each one links its source.
- Model
- Claude Haiku 4.5, by Anthropic
Sending what families signed up for
Every email here is assembled by fixed rules from the data above.
- Automation · no AI
Sends each family's weekly update
Builds one email per family from the schools they follow: the week's results, the next fixture and recent headlines for those schools. Assembled by fixed rules from the data above; a school with no news is shown as having none.
- How often
- Every Monday
- Last finished run
- Oct 5, 2026
- When it is unsure, or fails
- Anyone who has unsubscribed is skipped before sending, and if that list cannot be read nothing is sent. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Sends the public newsletter
Sends the weekly issue to the public list. A segment with nothing new that week is skipped, and every email carries one-click unsubscribe.
- How often
- Every Monday
- Last finished run
- Oct 5, 2026
- When it is unsure, or fails
- Opt-outs are a hard gate. After a send, the founder is sent a copy of what went out. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Sends follow-up reminders
Works out each family's reminders with the same rules the Tracker screen uses, and emails the ones that are newly due. At most one email a day, and each reminder once.
- How often
- Every day
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- If the record of what was already sent cannot be read, nothing goes out until it can. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Sends the emails that follow a sign-up
Sends a short series after a free sign-up, one note to finish an unfinished profile, one note after a checkout that was not completed, and one before a Season Pass ends. Each is sent once.
- How often
- Every day
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- Each email is recorded before it is sent, so it cannot go twice. If a due run is missing or fails, the daily watchdog emails the founder.
Checking the work
The jobs that watch the other jobs.
- Automation · no AI
Checks that every job ran
After the day's jobs have had their turn, it reads the job ledger, flags any job that did not finish or recorded a failure, re-syncs billing with the payment provider, and sends the email and AI services a test request to confirm they answer.
- How often
- Every day
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- Each problem it finds is emailed to the founder once that day. If the watchdog itself crashed or never ran, the next day's run says so.
- Automation · no AI
Measures how accurate the data is
Takes a sample of programs, re-reads each one's roster and staff page, and compares roster size, three random players' class year and position, the head coach's name and address, and the season with what the site holds. The result is an accuracy rate per field.
- How often
- The 20th of every month
- Last finished run
- No finished run on record yet. It runs on the schedule shown.
- When it is unsure, or fails
- It writes its measurements and nothing else. From the 21st, a month with no measurement is emailed to the founder.
- Automation · no AI
Tests every change to the site
Runs the type check and the whole test suite on every change to the code, including the tests that hold the published figures, the contact gate and this page to the truth.
- How often
- On every change
- Runs
- With every change. It keeps no record this page can read.
- When it is unsure, or fails
- A failed run shows as a red mark on the change. It does not block a deploy by itself.
The business side
Two more jobs, listed so the count on this page is complete.
- Automation · no AI
Counts the day's subscriptions
Writes one row a day: how many families are subscribed and what that adds up to. Internal accounts are left out.
- How often
- Every day
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- A failed count is recorded as a failure, not as a run. If a due run is missing or fails, the daily watchdog emails the founder.
- Automation · no AI
Rebuilds the list of clubs to call
Rebuilds the founder's list of club directors to phone. It never changes the outcome recorded for a call that was made.
- How often
- Every day
- Last finished run
- Oct 6, 2026
- When it is unsure, or fails
- If a due run is missing or fails, the daily watchdog emails the founder.
Software does the work.One founder answers for it.
No person reads every row or every answer. What keeps the data honest is in the jobs themselves: a doubtful result is held or dropped instead of published, each school’s roster carries the day it was read, and a daily watchdog emails the founder when a job it tracks is late or has failed.
If something on a school’s page looks wrong, email Ricky Ybarra at ricky.ybarra@getstartlineapps.com.