Football Intelligence Platform

An end-to-end data engineering project powering Krogvad Analytics Hub — football intelligence for the whole world. The platform builds a dedicated analytics home for every league, each with its own identity, all on one shared data warehouse and refreshed every night. It runs the full modern-data-stack pipeline from raw API to live dashboard.

Live → Krogvad Analytics Hub — the front door to every league site, all built from this pipeline.

---

Architecture

Sportmonks API   Highlightly API      Groq LLM      Poisson prediction model
       │                │                 │                    │
       └────────┬───────┴─────────────────┴────────────────────┘
                ▼
  Bronze layer        Raw JSON stored in MotherDuck (one table per endpoint,
                      per provider)
                       + LLM-generated match discussion rows
                       + pre-kickoff match predictions (frozen, never revised)
       │
       ▼
  Silver layer        Cleaned, typed, structured relational tables, each one
                      unioning every provider that carries the entity   (dbt)
       │
       ▼
  Gold layer          Kimball star schema  ─────────────────────────────┐
                      (dims + fct_team_matches                          │
                           + fct_player_appearances                     │
                           + fct_match_events                           │
                           + fct_match_predictions                      │
                           + fct_match_discussions                      │
                           + fct_team_transfers)  (dbt)                 │
                            │                                           ▼
                            │                             Evidence.dev dashboards
                            ▼                             (one app per league, plus the hub)
                  Public MotherDuck share                  deployed on Vercel
                  (read-only copy of gold)

The nightly GitHub Actions pipeline runs the four bronze producers in parallel (two API ingestions, LLM discussions, match predictions), then silver and gold sequentially with data-quality tests, then refreshes the public share and triggers a Vercel rebuild — so every site reflects last night's data.

---

Tech stack

Layer	Tool
Data sources	Sportmonks REST API · Highlightly REST API
Data warehouse	MotherDuck (DuckDB cloud)
Ingestion	Python (ingestion/sportmonks/, ingestion/highlightly/, ingestion/groq/, ingestion/datascience/)
Match predictions	In-house Poisson goals model (ingestion/datascience/predict_match_outcomes.py)
Transformations	dbt-duckdb (dbt/)
Orchestration	GitHub Actions (nightly + manual triggers)
BI / Dashboard	Evidence.dev (one app per league site, plus the hub)
Hosting	Vercel

---

Data model

The gold layer follows Kimball dimensional modelling. Six fact tables cover six business processes:

• fct_team_matches — one row per team per match (each fixture produces two rows, one per side); team-level stats, results, tactical data, extra-time and shootout goals, and expected goals where the provider carries them
• fct_player_appearances — one row per player per match; individual performance stats and ratings (~65 measures)
• fct_match_events — one row per in-match event (goal, card, substitution, VAR decision); event type/sub-type, minute on the match clock, running scoreline after the event, and per-group sequence — powering the Match Analysis timeline, league/team event-timing charts, team game-state (time leading/level/trailing) analysis, and referee VAR intelligence. Reloaded per fixture, since the provider rewrites a match's whole event list when VAR overturns a decision
• fct_match_predictions — one row per team per predicted fixture (mirroring fct_team_matches grain); pre-kickoff win/draw/loss probabilities, expected goals and expected points from the Poisson model, frozen three hours before kickoff and never revised — powering the Prediction Module page
• fct_match_discussions — one row per match per persona; LLM-generated fan discussion comments (via Groq) powering the Fan Forum on the Match Analysis page
• fct_team_transfers — one row per club per transfer; incoming/outgoing moves with fee, type, status, transfer partner, and player, powering the Transfer Intelligence page

erDiagram
    fct_team_matches {
        int date_sk FK
        int time_sk FK
        int match_sk FK
        int team_sk FK
        int opponent_team_sk FK
        int league_sk FK
        int stadium_sk FK
        int referee_sk FK
        int coach_sk FK
        int formation_sk FK
        int team_side_sk FK
        int match_result_sk FK
        int goals_scored
        int goals_conceded
        int goals_ht_scored
        int goals_ht_conceded
        int goals_scored_extra_time
        int goals_scored_penalty_shootout
        decimal expected_goals
        decimal ball_possession_pct
        int corner_kicks
        int yellow_cards
        int red_cards
        int saves
        int fouls
        int offsides
        int points_earned
    }

    fct_player_appearances {
        int date_sk FK
        int time_sk FK
        int match_sk FK
        int player_sk FK
        int team_sk FK
        int opponent_team_sk FK
        int league_sk FK
        int stadium_sk FK
        int referee_sk FK
        int coach_sk FK
        int formation_sk FK
        int position_sk FK
        int team_side_sk FK
        int match_result_sk FK
        int appearance_type_sk FK
        int minutes_played
        int goals_scored
        int assists
        int shots_total
        int shots_on_target
        int passes_total
        int passes_accurate
        int key_passes
        int tackles
        int interceptions
        int clearances
        int saves
        int yellow_cards
        int red_cards
        int fouls_committed
        decimal rating
    }

    fct_match_events {
        int date_sk FK
        int time_sk FK
        int match_sk FK
        int league_sk FK
        int team_sk FK
        int opponent_team_sk FK
        int player_sk FK
        int referee_sk FK
        int stadium_sk FK
        int team_side_sk FK
        int match_event_type_sk FK
        int match_minute_sk FK
        int event_group_sequence
        int home_score_after_event
        int away_score_after_event
        int event_count
    }

    dim_date {
        int date_sk PK
        date date
        int year
        int month
        varchar month_name
        int quarter
        int week_number
        int day_of_week
        varchar day_name
        boolean is_weekend
        varchar season
        boolean is_current_season
    }

    dim_time {
        int time_sk PK
        int hour
        varchar period_of_day
    }

    dim_match {
        int match_sk PK
        int match_id
        varchar match_round_type
        varchar match_round_name
        int match_round_number
        varchar match_name
        varchar match_short_name
        varchar match_type
        varchar match_status
        varchar match_result
        varchar kick_off_time
        varchar _source
    }

    dim_match_event_type {
        int match_event_type_sk PK
        varchar event_group
        varchar event_type_name
        varchar event_sub_type_name
        varchar event_type_code
        varchar event_sub_type_code
    }

    dim_match_minute {
        int match_minute_sk PK
        varchar minute_label
        varchar period_name
        int minute_of_match
        int stoppage_offset
        varchar minute_type
        varchar minute_bucket
        int minute_bucket_sort
    }

    dim_team {
        int team_sk PK
        int team_id
        varchar team_name
        varchar team_code
        varchar team_short_name
        varchar team_country
        int team_founded_year
        varchar team_logo
        varchar team_venue_name
        varchar team_venue_city
        int team_venue_capacity
        varchar _source
    }

    dim_opponent_team {
        int opponent_team_sk PK
        int opponent_team_id
        varchar opponent_team_name
        varchar opponent_team_code
        varchar opponent_team_short_name
        varchar opponent_team_country
        varchar opponent_team_logo
        varchar _source
    }

    dim_league {
        int league_sk PK
        int league_id
        varchar league_name
        varchar league_type
        varchar league_country
        varchar league_country_code
        varchar league_logo
        varchar league_country_flag
        varchar league_short_code
        boolean league_is_active
        varchar _source
    }

    dim_stadium {
        int stadium_sk PK
        int stadium_id
        varchar stadium_name
        varchar stadium_city
        varchar stadium_country
        varchar stadium_surface
        int stadium_capacity
        decimal stadium_latitude
        decimal stadium_longitude
        varchar stadium_key
        varchar _source
    }

    dim_referee {
        int referee_sk PK
        int referee_id
        varchar referee_common_name
        varchar referee_firstname
        varchar referee_lastname
        varchar referee_nationality
        varchar referee_key
        varchar _source
    }

    dim_coach {
        int coach_sk PK
        int coach_id
        varchar coach_display_name
        varchar coach_nationality
        varchar _source
    }

    dim_formation {
        int formation_sk PK
        varchar formation
    }

    dim_player {
        int player_sk PK
        int player_id
        varchar player_name
        varchar player_nationality
        date player_birth_date
        varchar player_position
        varchar player_detailed_position
        int player_height
        int player_weight
        varchar player_photo
        varchar _source
    }

    dim_position {
        int position_sk PK
        varchar position_name
        varchar position_short_code
        varchar position_group
    }

    dim_team_side {
        int team_side_sk PK
        varchar team_side
    }

    dim_match_result {
        int match_result_sk PK
        varchar match_result
    }

    dim_appearance_type {
        int appearance_type_sk PK
        varchar appearance_type
    }

    dim_persona {
        int persona_sk PK
        varchar persona_name
        int sort_order
        boolean is_active
        varchar bio
    }

    fct_match_discussions {
        int match_sk FK
        int persona_sk FK
        int date_sk FK
        int league_sk FK
        varchar message
    }

    dim_transfer_type {
        int transfer_type_sk PK
        varchar transfer_type_name
        varchar transfer_direction
    }

    dim_transfer_status {
        int transfer_status_sk PK
        varchar transfer_status
    }

    dim_transfer_partner_team {
        int transfer_partner_team_sk PK
        int transfer_partner_team_id
        varchar transfer_partner_team_name
        varchar transfer_partner_team_country
        varchar transfer_partner_team_logo
        varchar _source
    }

    fct_team_transfers {
        int transfer_id
        int date_sk FK
        int team_sk FK
        int transfer_partner_team_sk FK
        int player_sk FK
        int transfer_type_sk FK
        int transfer_status_sk FK
        int transfer_count
        int transfer_fee_eur
    }

    fct_match_predictions {
        int date_sk FK
        int match_sk FK
        int team_sk FK
        int opponent_team_sk FK
        int league_sk FK
        int team_side_sk FK
        decimal win_probability
        decimal draw_probability
        decimal loss_probability
        decimal predicted_goals_scored
        decimal predicted_goals_conceded
        decimal predicted_points
        varchar model_version
        timestamp predicted_at
    }

    fct_team_matches }o--|| dim_date : "date_sk"
    fct_team_matches }o--|| dim_time : "time_sk"
    fct_team_matches }o--|| dim_match : "match_sk"
    fct_team_matches }o--|| dim_team : "team_sk"
    fct_team_matches }o--|| dim_opponent_team : "opponent_team_sk"
    fct_team_matches }o--|| dim_league : "league_sk"
    fct_team_matches }o--|| dim_stadium : "stadium_sk"
    fct_team_matches }o--|| dim_referee : "referee_sk"
    fct_team_matches }o--|| dim_coach : "coach_sk"
    fct_team_matches }o--|| dim_formation : "formation_sk"
    fct_team_matches }o--|| dim_team_side : "team_side_sk"
    fct_team_matches }o--|| dim_match_result : "match_result_sk"
    fct_player_appearances }o--|| dim_date : "date_sk"
    fct_player_appearances }o--|| dim_time : "time_sk"
    fct_player_appearances }o--|| dim_match : "match_sk"
    fct_player_appearances }o--|| dim_player : "player_sk"
    fct_player_appearances }o--|| dim_team : "team_sk"
    fct_player_appearances }o--|| dim_opponent_team : "opponent_team_sk"
    fct_player_appearances }o--|| dim_league : "league_sk"
    fct_player_appearances }o--|| dim_stadium : "stadium_sk"
    fct_player_appearances }o--|| dim_referee : "referee_sk"
    fct_player_appearances }o--|| dim_coach : "coach_sk"
    fct_player_appearances }o--|| dim_formation : "formation_sk"
    fct_player_appearances }o--|| dim_position : "position_sk"
    fct_player_appearances }o--|| dim_team_side : "team_side_sk"
    fct_player_appearances }o--|| dim_match_result : "match_result_sk"
    fct_player_appearances }o--|| dim_appearance_type : "appearance_type_sk"
    fct_match_events }o--|| dim_date : "date_sk"
    fct_match_events }o--|| dim_time : "time_sk"
    fct_match_events }o--|| dim_match : "match_sk"
    fct_match_events }o--|| dim_league : "league_sk"
    fct_match_events }o--|| dim_team : "team_sk"
    fct_match_events }o--|| dim_opponent_team : "opponent_team_sk"
    fct_match_events }o--|| dim_player : "player_sk"
    fct_match_events }o--|| dim_referee : "referee_sk"
    fct_match_events }o--|| dim_stadium : "stadium_sk"
    fct_match_events }o--|| dim_team_side : "team_side_sk"
    fct_match_events }o--|| dim_match_event_type : "match_event_type_sk"
    fct_match_events }o--|| dim_match_minute : "match_minute_sk"
    fct_match_discussions }o--|| dim_match : "match_sk"
    fct_match_discussions }o--|| dim_persona : "persona_sk"
    fct_match_discussions }o--|| dim_date : "date_sk"
    fct_match_discussions }o--|| dim_league : "league_sk"
    fct_team_transfers }o--|| dim_date : "date_sk"
    fct_team_transfers }o--|| dim_team : "team_sk"
    fct_team_transfers }o--|| dim_transfer_partner_team : "transfer_partner_team_sk"
    fct_team_transfers }o--|| dim_player : "player_sk"
    fct_team_transfers }o--|| dim_transfer_type : "transfer_type_sk"
    fct_team_transfers }o--|| dim_transfer_status : "transfer_status_sk"
    fct_match_predictions }o--|| dim_date : "date_sk"
    fct_match_predictions }o--|| dim_match : "match_sk"
    fct_match_predictions }o--|| dim_team : "team_sk"
    fct_match_predictions }o--|| dim_opponent_team : "opponent_team_sk"
    fct_match_predictions }o--|| dim_league : "league_sk"
    fct_match_predictions }o--|| dim_team_side : "team_side_sk"

Measures are abridged above: fct_team_matches carries ~50 team measures and fct_player_appearances ~65 player measures in full.

Dimensional model bus matrix

The bus matrix shows which dimensions are conformed (shared) across business processes — the foundation of Kimball integration.

BUSINESS PROCESSES →	Team Match Performance	Player Appearance	Match Events	Match Prediction	Match Discussion	Team Transfers
COMMON DIMENSIONS ↓
Date	X	X	X	X	X	X
Time of Day	X	X	X
Match	X	X	X	X	X
Team	X	X	X	X		X
Opponent Team	X	X	X	X
League	X	X	X	X	X
Stadium / Venue	X	X	X
Referee	X	X	X
Coach	X	X
Formation

All dimension surrogate keys are stable across runs — new records get new SKs, existing records keep theirs. Sentinel rows (-1 Unknown, -2 Not Applicable) handle missing lookups, with all VARCHAR attributes filled with descriptive defaults (e.g. 'Unknown Stadium Country'). Every dim_* table therefore has two more rows than it has real entities.

Because gold is the product — every mart reads it, and surrogate keys live in incremental dimensions that must never be rebuilt — gold changes ship with evidence from scripts/gold_verify.py: snapshot prod gold, run the change locally, diff row counts, key stability and column values before it reaches a PR.

---

Dashboards

Every site is its own Evidence.dev app. They share the warehouse, the component library and the design language, but each league site has its own identity (country-coloured logo, hero and accent palette).

Hub — the group front door: a spinning globe highlighting covered countries, the story of the group, live group-wide counters sourced from the dimensions, and a shelf linking every league site.

League sites — each ships the same core page set, plus a persistent side pane (hideable) and a mobile bottom bar for navigation, an installable PWA manifest, and a shared footer showing data freshness.

Page	Description
Home	Season KPIs, current leader, and navigation
Standings	Full table plus whatever split, group or playoff structure the league uses
Match Results	Results by round with scorelines and Players of the Week
Match Analysis	Per-match deep dive: formation-true lineup pitch with player ratings and stats, match timeline, team stats, head-to-head, and the LLM-generated Fan Forum
Upcoming Fixtures	Next fixtures with head-to-head history, last-5 form guide, and local kick-off times
Upcoming Match Analysis	Pre-match preview: head-to-head record, form, and the model's view with expected goals
Prediction Module	The match model's track record: cumulative points race (actual & projected, drawn even before a season starts), upcoming predictions with probabilities and expected goals, accuracy by round, and the full prediction history — every prediction frozen before kickoff
League Intelligence	Season awards podium, standings race, cross-team landscape and radar benchmarks, per-domain team rankings, and "The Rhythm of a Match" — when goals and cards land across the match clock league-wide
Team Intelligence	Per-team KPIs, form, performance vs previous season, shooting and possession breakdowns, event-timing profile (when this team plays best / concedes), and game-state analysis (time spent leading, level and trailing, plus comebacks)
About	Project background, stack overview, and the build journey
Data Glossary	Definitions of all metrics and KPIs used across the dashboard

Sites whose provider history runs deep enough additionally ship these pages:

Page	Description
Referee Intelligence	Cards and fouls by referee, strictness rankings, VAR decisions (who VAR ruled against), and per-match discipline logs
Stadium Intelligence	Interactive stadium map, fortress rankings, and home-advantage stats
Player Intelligence	League top-player podium by any measure, individual player deep dive with profile, characteristics radar, performance timeline, and match log
Transfer Intelligence	Club transfer market: spend KPIs, record signing & sale, transfer volume and count by team, market trend over time, and a searchable transfer ledger — filterable by year, window, team, direction, type, status and fee disclosure

Marts are many small, purpose-built SQL sources rather than a few shared ones: each is pre-aggregated at build time into a parquet cache, so the browser does no aggregation work.

---

Project structure

.
├── ingestion/
│   ├── sportmonks/             # Bronze: pull from Sportmonks API → MotherDuck
│   │   ├── run.py              # Ingestion runner (incremental + full load)
│   │   ├── engine.py           # Metadata-driven fetch engine
│   │   ├── api.py              # Sportmonks API client
│   │   ├── db.py               # MotherDuck connection
│   │   └── config.py           # Endpoint manifest + env vars
│   ├── highlightly/            # Bronze: pull La Liga / Liga MX / Süper Lig from Highlightly API
│   │   ├── run.py              # Ingestion runner (rolling window / backfill)
│   │   ├── engine.py           # List + detail passes, budget-aware
│   │   ├── api.py              # Highlightly API client (daily-quota steering)
│   │   ├── db.py               # Bronze tables + resumable-work queries
│   │   └── config.py           # Leagues, seasons, budget constants
│   ├── groq/                   # Bronze: LLM match discussion generation
│   │   └── generate_round_discussions.py  # Groq API → fct_match_discussions
│   └── datascience/            # Bronze: pre-kickoff match predictions
│       ├── predict_match_outcomes.py      # Poisson goals model → win/draw/loss probs
│       └── README.md           # Prediction contract (freeze rules, schema)
│
├── dbt/                        # Silver + Gold transformations (dbt-duckdb)
│   ├── models/
│   │   ├── silver/             # 39 models: bronze JSON → structured tables,
│   │   │                       #   unioned across providers, keyed on (id, _source)
│   │   └── gold/
│   │       ├── dims/           # 21 dim_* models (Kimball dims)
│   │       ├── fct_team_matches.sql
│   │       ├── fct_player_appearances.sql
│   │       ├── fct_match_events.sql
│   │       ├── fct_match_predictions.sql
│   │       ├── fct_match_discussions.sql
│   │       └── fct_team_transfers.sql
│   ├── seeds/                  # Display names, provider stat-code maps,
│   │                           #   and curated data-correction overrides
│   ├── tests/                  # Custom SQL DQ assertions (silver + gold)
│   ├── macros/                 # League-agnostic helpers (timezone, match scope,
│   │                           #   round-type conformance, incremental filters)
│   └── dbt_project.yml
│
├── dashboards/                 # Evidence.dev BI apps (one per site)
│   ├── hub/                    # Krogvad Analytics Hub — group front door
│   │   ├── components/         #   Globe.svelte + generated globe-data.js, side nav
│   │   ├── scripts/            #   sync_league_logos.py
│   │   └── sources/hub/        #   League summary, group counters, freshness
│   ├── superligaen/            # Danish Superliga site (full page set)
│   │   ├── pages/              #   One .md file per dashboard page
│   │   ├── sources/            #   SQL marts queried at build time (parquet cache)
│   │   └── components/         #   Shared Svelte components (lineup pitch, timeline,
│   │                           #     radar, side nav, bottom nav, footer …)
│   ├── scotland/               # Scottish Premiership site (full page set)
│   ├── ligamx/                 # Liga MX site (core page set)
│   ├── turkey/                 # Süper Lig site (core page set)
│   └── spain/                  # La Liga site (core page set)
│
├── scripts/
│   ├── push_to_prod.py         # Push local DuckDB → MotherDuck (schema-selective)
│   ├── pull_from_prod.py       # Pull MotherDuck → local DuckDB
│   ├── gold_verify.py          # Snapshot/diff gold to prove a change did what it should
│   ├── refresh_share.py        # Mirror prod gold → the public share database
│   └── generate_globe_data.py  # Regenerate the hub globe geometry (run on roster change)
│
├── .github/workflows/
│   ├── master.yml              # Nightly: bronze (2 APIs + discussions + predictions
│   │                           #   in parallel) → silver → gold → DQ → share → deploy
│   ├── ci.yml                  # PR validation: Python syntax + dbt compile
│   ├── bronze.yml              # Manual bronze-only run (Sportmonks)
│   ├── highlightly.yml         # Manual bronze-only run (Highlightly, budget-aware)
│   ├── silver.yml              # Manual silver-only run (dbt)
│   ├── gold.yml                # Manual gold-only run (dbt)
│   ├── discussions.yml         # Manual LLM discussion generation
│   ├── predictions.yml         # Manual match prediction run
│   ├── dq.yml                  # Manual DQ test run
│   ├── share.yml               # Manual public-share refresh
│   └── vercel.yml              # Manual Vercel deploy trigger
│
└── requirements.txt

---

Nightly pipeline

master.yml runs at 02:37 UTC, after the day is over, and is also dispatchable with a full-load flag.

Stage	Jobs	Notes
Bronze	4 jobs in parallel	Two API ingestions plus LLM discussions and match predictions. The second API runs on a free 100-calls/day plan: it steers by the quota header, stops cleanly when the budget runs out and resumes the next night
Silver	1 job	Blocks only on the primary ingest — a failed discussion, prediction or secondary-provider run shows red but doesn't stop the pipeline; that data just stays stale
Gold	1 job	Incremental by default (INCREMENTAL_DAYS_BACK=7), full-refresh on demand
DQ	1 job	dbt test + dbt source freshness; a failure blocks publish
Share	1 job	Rebuilds the public share from prod gold
Publish	1 job	Force-pushes publish_dashboard/vercel, which is the only branch Vercel builds from

Each manual workflow mirrors exactly one nightly job, so any stage can be re-run in isolation against either database.

---

Public data share

Gold is mirrored nightly into a separate MotherDuck database (superligaen_public) and handed out read-only. MotherDuck shares are database-wide, so gold gets its own database rather than exposing bronze and silver alongside it; views are materialised as tables, since a view carries a reference to the database it was built in that a recipient cannot resolve. Objects are discovered from the catalog, so a new gold model joins the share — and therefore becomes public — the night after it ships.

---

Environments

Environment	MotherDuck database	dbt target	Triggered by
Dev	superligaen_dev (local DuckDB)	dev	Local / feature branches
Prod	superligaen	prod	GitHub Actions (main)
Public share	superligaen_public (gold only)	—	Nightly, after DQ passes

Dev runs against a local superligaen_dev.duckdb file. Use scripts/push_to_prod.py to push local data to MotherDuck dev for dashboard testing.

---

Local setup

# 1. Clone and create a feature branch
git clone https://github.com/SaUgKi1773/data-engineering-demo.git
git checkout -b dev/<your-feature>

# 2. Create virtual environment
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# 3. Configure environment
cp .env.example .env
# Fill in MOTHERDUCK_TOKEN, SPORTMONKS_API_KEY, HIGHLIGHTLY_API_KEY, GROQ_API_KEY
# and TARGET_DB (superligaen or superligaen_dev)

# 4. Run layers against local dev
python ingestion/sportmonks/run.py
python ingestion/highlightly/run.py --days-back 7
cd dbt
dbt seed --target dev
dbt run --select silver.* --target dev
dbt run --select gold.* --target dev

# 5. Push to MotherDuck dev for dashboard testing
cd ..
python scripts/push_to_prod.py --db superligaen_dev --schema silver gold

# 6. Run a dashboard locally (any folder under dashboards/)
cd dashboards/superligaen
npm install
npm run sources   # regenerates parquet cache from MotherDuck
npm run dev
# → http://localhost:3000

Note: dropdowns and other option lists are populated from the built parquet cache, so verify them with npm run build && npm run preview — dev mode can show stale options.

---

GitHub Actions secrets

Secret	Description
MOTHERDUCK_TOKEN	MotherDuck service token (read-write)
MOTHERDUCK_TOKEN_READONLY	MotherDuck read-only token (dashboard build)
SPORTMONKS_API_KEY	Sportmonks API key
HIGHLIGHTLY_API_KEY	Highlightly API key
GROQ_API_KEY	Groq API key (LLM match discussion generation)