Metrics

Play-by-Play video tool: the signature feature

Filter every WNBA or NBA boxscore action into a video playlist. Season · Team · Player · Focus (player plays, plays drawn, plays caused) · Play type (assists, steals, threes, turnovers, blocks, rebounds, fouls, free throws) · Advanced filters (opposing team, game segment, score margin, period).

Result: a queued video reel with matchup, period, and play description under each clip. 60 seconds to build a Jessica Shepard assist reel. 90 seconds to export a Kelsey Plum pull-up highlight for a Twitter thread or a Youtube video.

League stats: Four Factors, six tabs, every filter

Selected league Teams ranked across the Four Factors: Offense (ORtg, eFG%, AST%, TOV%, ORB%, FT rate) and Defense (DRtg, opp eFG%, opp AST%, opp TOV%, DRB%, opp FT rate). Every cell carries a colored percentile rank.

Toggle Adjusted to switch values into deltas from league average. Filter by date range, regular season vs playoffs, or limit to games against Top 5 / Bottom 5. Standings panel adds Top 5 / Bot 5 records — the kind of split that tells you whether a hot start is real or padded by easy matchups. Six tabs: Basics · Overall · Net · Offense · Defense · Shooting · Opp Shooting.

Team pages — lineups, trends, roster stats

Every 5-player combination the team has fielded, with minutes together, GP, all the shooting splits, plus-minus — filterable by minimum offensive possessions. Season sparklines (pick any stat, see per-game evolution for every player overlapped with the league average). Minute-by-minute court presence heatmap partitioned by quarter, with color intensity as usage indicator. Roster payroll, contract years, cap percentage.

Player pages — radar, chemistry, shot charts, game log

Five seasons stacked vertically with career row. 6-axis radar (Scoring · Playmaking · Rebounding · Defense · Shooting · Discipline) against position-bucket peers. Lineups featuring this player — best and worst NetRtg chemistry by teammate. Shot charts (shots taken + shots created via assists). Assist network (outgoing and incoming). Full career game log, every row linkable to the game page. Season trends sparklines with rolling window (3 / 5 / 10 games).

Games — calendar, score progression, advanced box scores

Month calendar with per-day game counts. Sidebar on each day: season series record, last 5 form line, tip-off times. On played days: final score, score progression line chart, top performers. Dedicated game page: by-quarter table, score arc, top performers, Four Factors mirrored across both teams, full box score with percentile ranks. Six tabs: Basics · Advanced · Shooting · Offense · Playmaking · Defense.

Players table — every player, every filter

Every WNBA player who's logged a minute this season. Two hundred players. One sortable table with Four Factors per player, percentile ranks baked in under each value. Filter by date range, playoffs, vs Top 5 / Bot 5. Compare modal, scatter chart export, glossary.

NCAAW — same engine, college side

Power 5 + Big East coverage (79 teams, ~1000 players per season). Independent ESPN crawler. UI patterns mirror the WNBA side — anyone fluent on Metrics WNBA reads NCAAW the same way. Players index, single-player pages with Basics / Advanced / Shooting / Profile / Game Log / Season Trends.

Reverse-engineering

Most of the data is available for anyone, but still protected. A standard Python request returns may return a 403. I built a system to gather it and cache it indefinitely. Now runs silently on every request.

PBP possession counter at 99.5% accuracy

Counting possessions from raw play-by-play events sounds simple. And-ones, technical fouls, clock resets, quarter boundaries, team rebounds — all edge cases that break naive implementations. The remaining 0.5% gap is in the official data itself.

30+ advanced metrics, implemented from scratch

True Shooting %, Effective FG%, Usage Rate, Win Shares, PER, Dean Oliver's Four Factors, on/off court ratings, shot zone analytics across 6 zones, league-adjusted stats. Every formula from the source papers and cross-referenced.

A full SPA with no framework

No React. No Vue. No bundler. Custom router, lazy-loaded page modules. The 8-tab analytics grid — sortable columns, sticky team rows, Per Game / Per 100 / Totals / Adjusted toggles — hand-coded, every line. Bundle size near zero. Full control.

A full admin panel, built into the product

Three-column dashboard, auto-refreshed every 30 seconds.

Stats — signups by tier (free / beta / pro), beta funnel conversion, 24h visit counts parsed from nginx access logs, activity feed.

Ops — beta invite queue (pending / approved / rejected), feedback inbox from testers, one-click invite dispatch via Brevo.

Health — live VPS card (CPU, RAM, disk as color-coded bars with sparklines over time), API service status, cache state (stacked bar: today / yesterday / this week / older), warmup history for the last 7 runs, 6 alert rules with on/off toggle and test email trigger, quick actions (trigger warmup, reload nginx, restart API — each requiring typed confirmation). All side-effectful actions are logged in an admin_actions table.

6 automated alert rules, email-triggered

A cron runs every 5 minutes and evaluates six rules against the metrics_history table (itself populated every 10 minutes via psutil + journalctl): memory above 90% for 10+ minutes, disk above 85%, CPU above 85% for 15+ minutes, 5xx errors above 20 per hour, workers below 1, fewer than 10 cache files written today. Anti-flap cooldown built in. Alert emails go out via Brevo the moment a rule fires. No third-party monitoring service needed.

MCP servers & agentic layer for maintenance

Routine admin tasks — subscription management, email list operations, deployment triggers, beta approvals — are handled through a set of MCP servers connected to the product's external services: Lemon Squeezy (payments), Brevo (transactional email + lists), GitLab (CI/CD pipeline).

This means an LLM can be instructed to approve a batch of beta applicants, trigger a deployment, resync a subscription, or send a newsletter — through natural language, without touching the codebase. The MCP layer sits alongside the runtime app as a maintenance agent, used exclusively for operator-side tasks. FastAPI talks directly to the external APIs at runtime; MCP handles the human-in-the-loop operations.

In practice: what used to be 15 minutes of manual API calls across three dashboards is now a single instruction. This is MCP in production — not as a developer tool, but as an operational layer that reduces maintenance overhead at scale.