RaceDelta

---

What it does

Most F1 sites give you tables. RaceDelta computes the differences between drivers and explains where they come from.

Compare

Any two drivers, four ways:

Mode	What it answers
Season	Head-to-head on the latest race, plus cumulative points across the year
Race	One Grand Prix, lap by lap
Track	Every visit to a single circuit
Conditions	Wet, dry, safety car, VSC, hot track (≥40 °C), cool track (<25 °C)

AI Lab

• Driver Rating — a 0–100 season score built from within-race z-scores across race pace, qualifying pace, consistency, tyre management, overtaking, defence and discipline.
• Driving DNA — a per-driver strength vector plus the most similar drivers by cosine similarity.
• Clustering — k-means, DBSCAN or agglomerative grouping of drivers by driving style, with generated labels ("Qualifying Specialists", etc.).

Strategy Lab

Tyre degradation per compound, stint analysis, pit-stop timing, and a pace model that predicts lap times from stint state. Uses XGBoost when available and falls back to a linear model when it isn't.

AI Race Analyst

Ask questions in plain language ("who's the strongest contender", "what's Antonelli's driving style"). The model answers through tool calls against the same endpoints the UI uses, and shows which tools it called. Runs on NVIDIA NIM or the Anthropic API; without a key it degrades to an offline intent mode.

Dashboard, Drivers, Teams

Live standings, race calendar, session results, per-round analytics with weather telemetry, tyre stint timelines and position-change charts.

---

Data

Source	Used for
FastF1	Lap timing, telemetry, weather, track status
OpenF1	Live driver roster, sessions, meetings, stints
Jolpica	Standings and schedule — the successor to the retired Ergast API

The deployed instance ships with a pre-ingested snapshot of the 2026 season: 12 races, 14,101 laps, 762 stints.

---

Running locally

Backend

cd race-delta-backend
python -m venv venv && venv/Scripts/activate     # Linux/macOS: source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env                             # then fill in the values you need
python app.py                                    # http://127.0.0.1:8000

The schema is created automatically at startup. To populate the database with a season:

python ingest_local.py

Frontend

cd race-delta-frontend
npm install
echo "VITE_API_BASE=http://127.0.0.1:8000" > .env.local
npm run dev                                      # http://localhost:5173

---

Configuration

Backend, all optional except in production:

Variable	Default	Notes
DATABASE_URL	sqlite:///racedelta.db	Postgres works too
CORS_ORIGINS	* in dev, empty in production	Comma-separated, no trailing slashes. Production allows nothing until you set it.
FLASK_ENV	development	production turns off debug and shortens standings computation
NVIDIA_API_KEY / ANTHROPIC_API_KEY	—	Enables the AI Analyst; NVIDIA takes precedence
ADMIN_TOKEN	—	Admin ingestion endpoints stay disabled unless set
RACEDELTA_CACHE_DIR	.cache	Persistent JSON cache with stale-while-revalidate
FASTF1_CACHE_DIR	./fastf1_cache	FastF1's on-disk session cache

Frontend:

Variable	Notes
VITE_API_BASE	Backend origin, no trailing slash. Baked in at build time — changing it requires a rebuild.

---

Deployment

Backend — Docker, on any container host. The included Dockerfile is sized for a 512 MB / 0.1 CPU instance: one gunicorn worker with eight threads, and the computed cache plus the season database baked into the image so a cold start serves real data without refetching.

Root Directory   race-delta-backend
Dockerfile Path  ./Dockerfile
Health Check     /api/health

Frontend — any static host. Root directory race-delta-frontend, Vite preset. vercel.json supplies the SPA rewrite that React Router needs for deep links.

Refresh the shipped season data with:

cd race-delta-backend && python ingest_local.py
git add -f data/racedelta.db && git commit -m "chore: refresh season data"

---

Tech stack

Frontend — React 18, Vite, Tailwind, React Router, Recharts, Plotly, Chart.js, Framer Motion

Backend — Flask, SQLAlchemy, FastF1, pandas, NumPy, scikit-learn, XGBoost (optional), gunicorn

---

Project layout

race-delta-backend/
  app/
    routes/          core, compare, lab, admin, openf1
    services/        f1_service, ingestor, feature_store, strategy_lab,
                     driver_intelligence, race_analyst, compare_lab, cache_store
  models.py          10 tables: races, race_sessions, laps, stints,
                     driver_race_features, drivers, constructors, ...
  scripts/           Jolpica/Ergast standings and team lookups
  data/              pre-ingested season database
  .cache/            precomputed API responses shipped with the image

race-delta-frontend/
  src/routes/        Home, Compare, Drivers, Teams, Lab, Strategy, Analyst, Race
  src/components/    Navbar, cards, charts, comparison tables
  src/api/client.js  single source of truth for API access

---

Notes and limits

• The deployed season data is a snapshot, not a live feed. Standings and the driver roster are live; Compare, Strategy and AI Lab read the shipped database until it's regenerated.
• Free-tier hosting sleeps after 15 minutes idle, so the first request after a quiet period takes ~50 s.
• /api/compare/track-map loads full telemetry and is memory-hungry; it needs more than 512 MB.