reviewr

Inspiration

Reviewr started from a simple frustration: reviewing gameplay VODs is valuable, but it is also slow, repetitive, and difficult to scale.

AI can help identify potential coaching moments, but generating more moments is not enough. Candidate queues can still contain vague explanations, unsupported conclusions, repeated observations, or moments that require so much editing that the coach may as well start from scratch.

We wanted to build a workflow where AI helps surface and challenge coaching insights without replacing the coach.

The guiding idea became:

AI should help separate signal from noise, while the human remains responsible for the final review.

What it does

Reviewr is a gameplay VOD review workspace for esports coaches, analysts, and players.

It supports a private review workflow where candidate gameplay moments can be inspected, edited, rejected, approved, and eventually published as a focused shared review.

For Build Week, we added a multimodal evidence-review layer.

For one bounded gameplay moment, Reviewr:

1. Loads an existing private coaching candidate.
2. Samples a small chronological sequence of gameplay frames.
3. Combines those frames with bounded structured evidence.
4. Asks GPT-5.6 to recommend Keep, Revise, or Reject.
5. Validates the response against a strict schema.
6. Displays the recommendation privately to the coach.
7. Applies nothing unless the coach explicitly chooses to act.

A recommendation may identify:

• Observations supported by the gameplay evidence
• Unsupported or overly confident claims
• Uncertain conclusions
• A more precise title or explanation
• A concrete coaching action
• The evidence identifiers used

GPT-5.6 cannot automatically save, reject, approve, or publish a moment. The coach retains control over every user-visible change.

Reviewr also includes deterministic replay using the same evidence-review contract, allowing the complete workflow to be tested reliably without requiring a live provider call.

How we built it

Reviewr is built with a Python backend and a React and TypeScript frontend.

The Build Week contribution was deliberately isolated from the larger existing Reviewr codebase.

The new backend feature includes:

• Strict request and response schemas
• Bounded evidence-packet construction
• Ordered frame sampling
• Frame-count and payload-size limits
• Candidate-revision hashing
• Stale-result protection
• A fixture and live provider interface
• Deterministic Keep, Revise, and Reject fixtures
• A multimodal OpenAI Responses API adapter
• An owner-authenticated evidence-review endpoint
• Public-payload non-leakage protections

The frontend includes:

• A focused evidence-review panel
• Clear deterministic versus live states
• Visual-evidence status
• Supported and unsupported claim displays
• Explicit Apply revision and Dismiss recommendation controls
• Stale-recommendation handling
• Preservation of private approval and publication states

The gameplay-backed demo uses a real Pokémon UNITE VOD. Reviewr selects only a short candidate window and samples a bounded ordered frame sequence rather than sending an entire VOD to the model.

Codex was used to:

• Audit the existing repository
• Identify the safest integration seam
• Separate pre-existing work from Build Week work
• Design the evidence-review contract
• Implement the backend and frontend feature
• Build deterministic fixtures and focused tests
• Diagnose runtime and media preparation
• Verify the browser workflow
• Create judge-facing setup and architecture documentation

Challenges we ran into

The largest challenge was working inside an existing AI-grown codebase that already contained large backend and frontend integration files.

Rather than attempting a risky full rewrite, we treated the existing application as legacy infrastructure and placed the new feature behind a small, readable module boundary.

Other challenges included:

• Separating pre-existing Reviewr work from the new Build Week contribution
• Preventing model recommendations from silently overwriting human edits
• Binding recommendations to the exact candidate revision
• Keeping model diagnostics private while allowing approved coaching content to be shared
• Supporting deterministic replay without pretending it was a live model response
• Sampling useful chronological gameplay frames while keeping payloads bounded
• Preparing a reproducible gameplay-backed demo without committing large video or frame assets
• Preserving a stable demo despite provider availability, API credits, and media-tool differences

We also had to resist expanding the scope into full-VOD multimodal analysis, analyzer rewrites, recommendation history, or a broad UI redesign.

Accomplishments that we're proud of

We are proud that the resulting feature remains human-controlled from beginning to end.

In particular, we:

• Isolated the judged contribution into a small set of purpose-specific modules
• Added multimodal review for bounded gameplay windows
• Created strict Keep, Revise, and Reject contracts
• Required evidence identifiers for model-supported claims
• Added deterministic replay using the same result schema as live mode
• Prevented recommendations from automatically changing approval or publication state
• Ensured private diagnostics do not appear in shared reviews
• Created a real gameplay-backed demo with coach-selected moments
• Maintained a clean, attributable commit history
• Added focused backend and frontend tests
• Preserved the existing Reviewr product rather than destabilizing it with a broad rewrite

The most important accomplishment is that GPT-5.6 acts as a disciplined second reviewer rather than an autonomous coach.

What we learned

We learned that the most valuable role for AI in a review workflow is not necessarily generating more content.

The harder and more useful problem is deciding:

• Whether a moment is actually supported
• Whether a conclusion is too confident
• Whether the coaching action is specific enough
• Whether the moment is worth keeping at all

We also learned that multimodal inputs are most manageable when they are bounded. Sending a small, ordered frame sequence for one known moment is easier to validate, cheaper to run, and more useful than pretending a model fully understands an entire VOD.

Another major lesson was architectural: an existing complicated product does not need to be completely rewritten to produce a readable new contribution. A narrow service boundary, strict schemas, explicit human controls, and focused tests can make one feature understandable even when the surrounding application remains large.

Finally, deterministic replay proved valuable not as a substitute for live AI, but as a reliable way to test UI behavior, state safety, schema validation, and judge workflows.

What's next for Reviewr

The next steps are to expand the evidence-review system carefully rather than simply increasing the amount of generated analysis.

Future work includes:

• Validating Keep, Revise, and Reject calibration across more games and coaching styles
• Evaluating more GPT-5.6 model tiers for quality, latency, and cost
• Improving game-specific evidence extraction
• Strengthening temporal reasoning across longer or more complex moments
• Adding better coach-facing evidence visualization
• Supporting team discussion around approved moments
• Measuring how much the reviewer reduces editing time
• Improving moment discovery while preserving the post-persistence evidence gate
• Adding collaboration and review-history features
• Extending Reviewr beyond esports into other video-based coaching and analysis workflows

The long-term goal is not autonomous coaching.

It is a trustworthy review workspace where AI handles more of the evidence organization and quality control, while humans retain authorship, judgment, and accountability.

Built With

• codex
• python
• typescript