Sputnik Digital Graphics

PigskinAI

A fantasy football brain built in the cloud. Data first, takes second.

The studio's deepest engineering mission to date: a full fantasy football intelligence platform on Google Cloud. It ingests the entire modern NFL data record into BigQuery, learns position-specific ranking formulas, explains every rank through a Gemini-powered analyst, and publishes cryptographically verifiable boards the public can read.

See the live rankings See the technology
Pigskin AI elite tier ranking board where each player's rank is explained by touches, target share, red-zone usage, and efficiency metrics
Scoring profiles 4
Boards per release 20
NFL data sources 15+
Warehouse BigQuery
Feed integrity SHA-256
01 The platform

Rankings that can show their work.

Most fantasy rankings are vibes. Pigskin AI is the opposite: deterministic formulas over a real data warehouse, with an AI analyst that explains every rank from the player's actual metrics.

The foundation is a BigQuery warehouse loaded from the modern NFL data record: play-by-play, Next Gen Stats, charting data, snap counts, injuries, depth charts, contracts, and drafts. Derived truth tables separate role quality from box-score noise, so the system knows the difference between a player who earned his points and one who lucked into them.

On top sit machine-learned, position-specific ranking formulas that are backtested before promotion, a Gemini-powered analyst that must query the warehouse before making a claim, and an admin studio on Cloud Run for research, review boards, and show production. Four scoring profiles ship per release: Standard, PPR, Half-PPR, and the GNG Keeper league's custom model.

  • NFL data warehouse
  • ML ranking formulas
  • AI analyst with tool calling
  • Gated release pipeline
  • Verified public feed
  • Show production tools
Pigskin AI tiered ranking board crediting Rankings by Pigskin AI, the brainchild of Sputnik Digital Graphics

The same release flows from BigQuery to a verified public feed to the boards fantasy players actually read.

Pigskin AI Public rankings: The consumer surface on TheGNG.us, fed daily
Public rankings The consumer surface on TheGNG.us, fed daily
Pigskin AI AI verdicts: Every rank explained from its real formula inputs
AI verdicts Every rank explained from its real formula inputs
Pigskin AI Tiered boards: Elite to front-line tiers with scores and projections
Tiered boards Elite to front-line tiers with scores and projections
Pigskin AI Verified feed: Immutable SHA-256 boards behind a versioned manifest
Verified feed Immutable SHA-256 boards behind a versioned manifest

Every system on the stack.

Pigskin AI is six interlocking systems: cloud infrastructure, data engineering, machine learning, the AI analyst, a gated production pipeline, and public distribution. All of it built in-house.

Sys 01

Cloud infrastructure

Everything runs serverless on Google Cloud: the admin studio, the scheduled jobs, the secrets, and the public feed. No servers to babysit.

  • Cloud Run + Cloud Run Jobs The Streamlit studio as a service, with long-running ingestion and ranking jobs split out of the request path
  • Cloud Build + Artifact Registry Dockerized builds published as immutable, explicitly tagged images. No deploying latest.
  • Cloud Scheduler Timed triggers for recurring ingestion and materialization jobs
  • Secret Manager + IAM hardening API keys out of code, least-privilege service accounts, and a public bucket that exposes objects only
  • Cloud Storage Artifact archives plus the dedicated public rankings bucket
Sys 02

Data engineering

A BigQuery football warehouse that turns raw NFL feeds into precomputed truth tables, so every downstream answer starts from role quality instead of box scores.

  • BigQuery warehouse Partitioned, clustered analytical marts as the single source of truth
  • nflverse ingestion Play-by-play, weekly metrics, NGS, FTN charting, snap counts, injuries, depth charts, contracts, and draft data via nflreadpy
  • Analytical truth tables Derived marts for player role quality, fraud detection, QB splits, and game environment
  • External context feeds Sleeper league snapshots, FantasyCalc market values, Reddit RSS, CollegeFootballData, and scouting imports
  • Idempotent ETL + validation Re-runnable pipelines with validation queries for every transformation
Sys 03

Machine learning & formulas

Rankings come from deterministic, position-specific formulas developed with BigQuery ML, backtested against history, and approved before they ever touch production.

  • BigQuery ML models Positional ranking architecture researched and trained inside the warehouse
  • Position-specific formulas Dedicated QB, RB, WR, and TE formulas tuned per scoring profile
  • Backtesting framework Historical validation with review boards and owner-approval packets before promotion
  • Model provenance Every output versioned by model run with source freshness, config version, and scoring profile recorded
Sys 04

The AI analyst layer

Pigskin is a Gemini-powered football analyst that must query the warehouse before it speaks, and admits failure instead of inventing numbers.

  • Google Gemini (google-genai) Custom analyst persona with strict system instructions and a show-ready voice
  • BigQuery function calling Manual tool calling so every analytical claim is backed by a real query
  • Hallucination guards Failed queries stop the answer; SQL self-repair fixes known schema mistakes
  • Vertex AI Search External context verification beyond the warehouse
  • Metric-backed verdicts Plain-language rank explanations generated from each player's actual formula inputs
Sys 05

Production ranking pipeline

Publishing a board is a gated, auditable release: dry runs first, explicit write gates, invariant checks, and a documented recovery path.

  • Staged promotion Formula, safety review, positional boards, unified Top 100, and publication as separate owned stages
  • Top 100 interleaver Cross-position board that provably preserves each position's internal order
  • Release invariants Contiguous ranks, normalized public scores, zero teamless players, and non-empty verdicts, or the release fails closed
  • Runbooks + board history Canonical release procedure, archived prior versions, and tested recovery steps
Sys 06

Public feed & show tools

The results ship two ways: a cryptographically verifiable public JSON feed, and content tooling built for fantasy football show production.

  • Content-addressed JSON feed Immutable SHA-256 named boards with a versioned manifest, served anonymously from Cloud Storage
  • TheGNG.us integration Daily hash-verified imports power the public rankings pages
  • Show segment tools Fraud Watch, Sleeper Watch, Player Profiles, Versus Finder, and Reddit Topic Scout
  • Trade + market analysis FantasyCalc values, age curves, and AI synthesis across multi-year windows
  • Script-mode output Show-ready analysis formatted for text-to-speech and voice acting
02 The pipeline

From play-by-play to published verdict.

A ranking release is treated like a software release: staged, gated, validated against hard invariants, and recoverable when something breaks.

Scheduled jobs ingest fresh NFL data into BigQuery. Formulas score their positions and go through dry runs before any write gate opens. A unified Top 100 interleaves the positional boards without reordering them, verdicts are generated from each player's real formula inputs, and the publisher fails closed unless every board passes every invariant. Only then do immutable, hash-named boards and a versioned manifest go public, where TheGNG.us verifies the bytes against the manifest before importing a single row.

  • Scheduled cloud ingestion
  • Backtest-gated formulas
  • Dry-run promotion gates
  • Invariant-checked releases
  • Immutable public boards
  • Hash-verified consumers
See it live on TheGNG.us

Have a hard data problem?

Pigskin AI proves the studio can take a project from raw data to warehouse to machine learning to a product people use every day. If you need that kind of depth, open a channel.

Start your build