Computer vision · Sports analytics

Smashpoint

A layered video-analysis system that turns Super Smash Bros. broadcasts into auditable timelines, events, and player-level match insights.

Role: Product architecture, computer vision, model evaluation, analytics, and dashboard engineering
Period: 2026
Status: Active prototype

Tournament match frame with detected timer, portraits, damage values, tags, and stock regions outlined — Auditable HUD extraction on a tournament broadcast frame.

6pipeline layers

1stable timeline contract

Humanposition confirmation

Challenge

What the work needed to resolve.

Recover reliable match state from compressed, overlaid tournament video while preserving enough evidence to audit every automated conclusion.

Approach

A system designed for scrutiny.

01
Separated ingest, signals, timeline fusion, events, analytics, and visualization behind a typed timeline contract.
02
Combined structured HUD extraction with precision-gated on-stage cues and explicit human review.
03
Built resumable VOD processing and a dashboard for discovery, set summaries, review, and broadcast views.

IngestSignalsTimelineEventsAnalyticsVisualization

Inside

What the system actually produces.

The figures below are rebuilt in the portfolio from synthetic series, so they show the shape of each output without carrying any real data out of the project.

From frames to a timelineSynthetic data

Per-player damage recovered from the HUD, with stock losses resolved as timeline events. Everything downstream reads this one contract.

Precision gate, then a humanSynthetic data

Events below the confidence gate are not silently dropped or silently accepted — they are queued for review, so the automated conclusion stays auditable.

Outcome

The prototype indexes games and matchups, derives stock-loss and damage summaries, and exposes uncertainty instead of hiding it behind a single score.

Continue exploring

Open the workspace Source

Next case studyNotezeit Media Lab