Inspiration
Before every match, coaches scrub hours of VODs to answer two questions: where does the opponent put their utility, and who holds the Operator? The official GRID feed already records every cast, every position and every kill. I wanted a tool that answers those questions in seconds, and that a coach can trust because every claim links to the rounds behind it.
What it does
Cirrus is a scouting app for Cloud9's VALORANT and Game Changers coaches. Pick a team, a map and a side, and it shows:
- Findings in a coach's language, with sample sizes. "Killjoy Turret at B Kitchen: 45 of 84 defense rounds, placed by cortezia, buy phase." "aspas opens on the Operator at B Yellow: 16 of 37 Operator rounds, first 20 seconds."
- Utility map. Repeated spots on a floor plan, sized by how many rounds use them, with a density heatmap. Filters cover utility type, round phase, side and period.
- Operator profiler. Buy rate, the primary holder, where the Operator opens (each round counted once), the share of sightings by round time (0–20 s, 20–40 s, 40 s+) labelled aggressive, on-site or passive, and kill sightlines from shooter to victim.
- Setups. A team's recurring trap and cam setups, with a predictability score shown against every team on the map.
- Counters tested across the league. Select a defensive spot to see whether attackers who hit it with utility early won more, across every round where any defense used that spot. Against a Vyse Shear at C Main on Lotus, attackers won 53.8% when they hit it early and 47% when they left it, across 939 rounds.
- Game plan. Those counters for a team's whole defense in one list: which setups are worth hitting early, which are better left alone, and which show no reliable edge. Each row opens the spot on the map with its numbers.
- Evidence and 2D replay. Each finding lists its rounds. One click replays a round in 2D straight from the feed.
- Self-scouting. The same analysis on Cloud9: their own turret sits at Mid Boiler in 76% of their Icebox defense rounds.
- A brief to take away.
python -m scout briefwrites the same findings as a short Markdown note for a prep document.
How I built it
- Extraction. A streaming parser reads all 857 VALORANT series files, about 160 GB once decompressed, without unpacking them to disk. It rebuilds every round: sides, buys, utility casts with the supplement's precise positions, kills, plants, and Operator possession tracked through buys, drops, pickups and deaths.
- Integrity gates. A map is analysed only if its rebuilt rounds add up to the official score. A round is analysed only if its result agrees with its sides.
- Original map art. Floor plans are traced into vectors from where players actually stood, so no Riot art is used and everything lines up with the data.
- Front end. React and TypeScript with a single SVG map. The same build runs against the local API or an aggregate-only static export for the hosted demo.
- Hosted demo without storing data. GitHub Actions fetches the official data for each run, processes it, runs the tests against it and deploys only aggregate statistics. The organizers confirmed this is allowed under the data license.
- Tests. 85 automated tests, including an exact recount of a headline number from the raw GRID files by separate code.
Methodology
Every number is defined so that anyone can recount it. The full write-up is docs/METHODOLOGY.md in the repository. In short:
- A spot's share is the number of rounds with a cast of that ability within 4.5 m of the spot's centre, divided by the rounds sampled. It is shown as "k of n rounds", never as a bare percentage.
- Setups as they stood. A device picked up and placed again in the buy phase counts once, where it finally stood. Charge counts are learned from the data.
- Operator opening area is where the holder spent most of the first 20 seconds, with each round counted once, so walking through an area does not count as holding it.
- Evidence strength uses a Wilson lower bound: strong needs 15 or more rounds and a lower bound of 40%, moderate needs 8 rounds and 20%, anything else is labelled thin. Thin findings are shown as thin, not hidden.
- Counters compare attacker win rates with a two-proportion z-test and are labelled clear signal, some signal or noise. The game plan sorts a team's repeated setups by that label and adds no thresholds of its own.
- What the data cannot say. The feed gives where a player stood when casting, not where the ability landed, and the app says so on every thrown ability. Maps with one position sample per second are flagged in every view that uses them.
JetBrains tools
I used Junie, JetBrains' coding agent, through the Junie CLI. Every task it was given is in the repository under docs/junie-tasks, in the order it ran, and docs/junie-case-study.md tells the story.
- Guidelines first. Junie read the codebase and wrote its own project guidelines (
.junie/AGENTS.md): the pipeline, the commands, the data-license rules and the analysis conventions a change must respect, such as the single place percentages are rounded. Junie loads that file on every later task. - Two features from written briefs. Junie built the scouting brief (
scout/brief.pyand thepython -m scout briefcommand) and the game plan (the classification inscout/analysis.py, the payload, the brief section and the section in the web app), with their tests. Each brief set hard rules, such as reading every number from the app's payload and inventing no new thresholds. - Review rounds, not blind trust. Junie's first version of the brief passed its own tests but used a 30-day window where the app uses six months, so it reported 23 rounds where the app reported 84. I found that by running it against the app, wrote the findings up as a new task, and Junie fixed them. For the game plan, a separate script re-derived every result on six real views and the analysis was right the first time; the review found presentation problems, which Junie fixed.
- Where a person was needed. Junie cannot see the page. Its game plan rows worked but looked like a different app, and I restyled them by hand after looking at screenshots.
Junie joined after the pipeline, the analysis and the web app were in place. Its work is the guidelines, the scouting brief and the game plan, and the six task files are the complete record of it.
Challenges I ran into
- A cutoff that hid setups. An early version dropped utility cast more than 50 seconds before the round went live. A recount from the raw files did not match, which exposed it: 28,465 casts had been lost, mostly setup utility in rounds with timeouts.
- Snapshots are partial. GRID sends full inventories only occasionally, and partial snapshots leave out a team's side, so Operator possession had to be tracked as a state machine.
- Smokes with no location. Smokes placed from the tactical map have no meaningful cast position, so they are listed, not drawn.
What I learned
Passing tests are not proof. Both the lost setups and the brief's wrong time window were found by recounting from a second source, not by the test suite. Writing the repository's conventions down first made the coding agent's work far more consistent.
What's next for Cirrus
- Estimate landing points for thrown utility by fitting per-ability throw models to damage events.
- VOD deep links once broadcast offsets are available.
- A next-round predictor from earlier rounds in the same series.
Match data is provided by GRID. VALORANT is a trademark of Riot Games, Inc. This project is not endorsed by Riot Games, GRID or Cloud9.