Federico Seco
MenuClose

01 / AI-enabled coaching product

Turn one finished match into one better decision next game.

A post-match coaching product where deterministic code extracts and validates the evidence around one bounded model-generated review.

Product brief

Product
LeagueLoop
Status
Validated locally
Target user
A competitive League of Legends player who wants one credible adjustment after a completed match, not a dashboard full of generic advice.
My role
Product boundary, deterministic evidence pipeline, model contract, evaluation design, privacy controls, report UX, and portfolio surface.

The problem

Match telemetry is abundant, but raw statistics do not explain which decision mattered, why it mattered, and what a player should try next.

Open coaching demo ↗All work →

01 / Workflow

The product, end to end.

Follow the product from first input to a reviewable outcome.

LeagueLoop product flow
  1. 01Select one completed match
  2. 02Extract candidate evidence
  3. 03Rank relevant moments
  4. 04Generate one bounded review
  5. 05Validate claims
  6. 06Read or export coaching report

02 / Product surface

The interface carries the evidence.

Owned, locally generated product assets—no client screenshots or borrowed logos.

LeagueLoop coaching review product overview
A report-first surface makes the coaching outcome—not the generation machinery—the product.
First page of a LeagueLoop coaching report
The printable review begins with a clear adjustment and the evidence supporting it.
Second page of a LeagueLoop coaching report
Deeper evidence remains scannable and player-facing.

03 / Decisions

Architecture with a reason.

Next.js · TypeScript · PostgreSQL · Structured model output · Vitest · PDF export

01

Code owns the evidence

Deterministic extraction and ranking surround the model so generated coaching cannot invent match facts unchecked.

02

One review, explicit cost

Paid generation is gated, observable, and never hidden inside a browse or refresh action.

03

Human coaching language

The player sees what happened, why it mattered, and what to try next; diagnostics remain secondary.

Failure paths

What can go wrong is part of the product.

  • Insufficient match evidence for a confident claim
  • Structured response fails validation
  • Provider timeout or paid-call refusal
  • Local retention or deletion cannot complete

Verification

What supports the claim.

  • Offline evaluation fixtures
  • Structural and semantic validators
  • Stored-corpus regression runs
  • Dedicated report PDF rendering review

Portfolio boundary

Complete enough to review. Honest enough to trust.