CS2 Analytics Platform

Full-stack analytics platform for Counter-Strike 2. Upload match demos, get win probability predictions, player ratings, error detection, and tactical analysis powered by trained ML models.

Source-available for portfolio review. Licensed under PolyForm Noncommercial 1.0.0 — free for personal, research, and educational use. Commercial use requires a separate license (contact: pedrom02.dev@gmail.com). Trained model checkpoints and the proprietary demo dataset are not included in this repository.

Tech Stack

Layer	Technology
Frontend	Next.js 16, React 19, TypeScript, Tailwind CSS, shadcn/ui, Recharts, D3.js
Backend	FastAPI, SQLAlchemy 2.0, Pydantic 2, Celery, Redis
ML	PyTorch (Mamba SSM), LightGBM, CatBoost, UMAP, SHAP
Database	PostgreSQL 16, Redis 7
Infra	Docker Compose (dev), Kubernetes manifests + Helm (scaffolded)

What Works

Core pipeline (functional):

• JWT auth (register/login/refresh/logout) with org isolation
• Demo upload to MinIO, parsing via Awpy, feature extraction
• Match detail with round stats, economy breakdown, win probability impacts
• Player stats with aggregated KD, ADR, rating, headshot %
• Scout reports (CRUD) with weakness/strength profiles
• Stripe billing (checkout, portal, webhook signature verification)
• Pro match data from HLTV (manual CLI ingestion, not automated)

ML models (5/6 trained):

Model	Framework	Metric	Status
Win Probability v2	LightGBM	AUC 0.904	Trained, real inference
Player Rating v1	CatBoost	R² 0.998	Trained, real inference
Player Archetypes v1	UMAP + HDBSCAN	8 clusters	Trained
Positioning v2	Mamba SSM	MAE 0.067	Trained
Timing v1	Mamba SSM	Acc 0.81	Trained
Strategy GNN v1	GraphSAGE	Coarse (6T/5CT)	Weak-supervision pipeline in ml-models/src/training/train_strategy_gnn.py (heuristic fallback still active until checkpoints are committed)

SHAP explainability for win_prob and player_rating via POST /api/v1/ml/explain.

Shipped in this iteration:

• Sentry integration (backend + frontend) + CSP / HSTS headers
• Dependency scanning (pip-audit + pnpm audit) in CI, Dependabot config
• Auth edge-case tests (expired/forged JWT, refresh reuse, concurrent rotation)
• Stripe webhook idempotency via Redis dedup
• Strategy GNN training pipeline (weak supervision, coarse taxonomy)
• ML feature-drift middleware + daily Celery job + Prometheus alert rules
• Real radar-asset loading in the replay canvas
• Coaching insights page (/matches/[id]/coaching)
• Terraform modules filled out (VPC, RDS, EKS + cert-manager/external-secrets/kube-prometheus-stack/loki)
• K8s NetworkPolicy, RBAC, External-Secrets, cert-manager ClusterIssuers
• Steam OpenID sign-in (/auth/steam/login + callback, "Sign in with Steam" button)
• Real-time SSE win-prob stream (/live/{match_id}/win-prob/sse)
• Public API with API keys (/public/*, /api-keys) + migration 008_add_api_keys
• Team overview endpoint + dashboard page
• k6 baseline load-test script + staging smoke-test + secret rotation script

Still partial:

• Replay: real Valve radar PNGs need to be dropped into packages/frontend/public/radars/
• GNN checkpoints are produced by the training script but not committed — run python -m src.training.train_strategy_gnn to populate
• Terraform has not yet been applied to a real AWS account

Project Structure

packages/
  frontend/           Next.js 16 App Router, shadcn/ui, TanStack Query
  backend/            FastAPI REST API, Alembic migrations (7), Celery
  ml-models/          Training scripts, model registry, SHAP API
  feature-engine/     Round/player/team feature extractors
  demo-parser/        CS2 .dem parsing via Awpy
  pro-demo-ingester/  HLTV scraper (Playwright) + demo downloader
infra/
  docker-compose.yml  PostgreSQL 16, Redis 7, MinIO
  k8s/                Kubernetes manifests (scaffolded)
  helm/               Helm chart (scaffolded)
  monitoring/         Prometheus, Grafana dashboards, Alertmanager configs
scripts/backup/       Postgres + model backup/restore scripts
docs/                 Architecture, deployment, contributing guides
.github/workflows/    CI: lint, test, build, e2e

Quick Start

git clone https://github.com/Pedrom2002/aics2.git
cd aics2
cp .env.example .env

# Infrastructure
docker compose -f infra/docker-compose.yml up -d

# Backend
cd packages/backend
uv sync --extra dev --extra test
alembic upgrade head
uvicorn src.main:app --reload --port 8000

# Frontend
cd packages/frontend
pnpm install
pnpm dev    # http://localhost:3000

Environment Variables

Variable	Description
DATABASE_URL	postgresql+asyncpg://user:pass@localhost:5432/cs2analytics
REDIS_URL	redis://localhost:6379/0
JWT_SECRET	Auth signing secret (min 64 chars)
STRIPE_SECRET_KEY	Stripe API key
STRIPE_WEBHOOK_SECRET	Stripe webhook verification
NEXT_PUBLIC_API_URL	Backend URL for frontend
CORS_ORIGINS	Allowed origins (comma-separated)

Full list in .env.example.

API

Router	Prefix	Status
auth	/api/v1/auth	Register, login, refresh, logout
demos	/api/v1/demos	Upload, list, parse, delete
players	/api/v1/players	Stats, archetypes
win-prob	/api/v1/win-prob	Round win probability + impacts
tactics	/api/v1/tactics	Strategy analysis (heuristic)
scout	/api/v1/scout	CRUD scouting reports
ml	/api/v1/ml	Model registry, SHAP explain
billing	/api/v1/billing	Stripe checkout, webhooks
pro	/api/v1/pro	Pro match metadata
health	/api/v1/health	DB + Redis probes
heatmap	/api/v1/heatmap	Stub
sse	/api/v1/demos/.../sse	Stub

Swagger docs at http://localhost:8000/docs (dev only).

Scripts

# Root (Turborepo)
pnpm dev / build / lint / test

# Backend
uv run ruff check src/ tests/
uv run pytest tests/ -v
alembic upgrade head

# Frontend
pnpm type-check
pnpm exec playwright test   # needs running backend

# ML training
cd packages/ml-models
python -m src.training.train_win_prob
python -m src.training.train_player_rating

# Pro demo ingestion (manual)
cd packages/pro-demo-ingester
python src/download_demos.py

CI

GitHub Actions on push/PR:

• Backend: ruff lint + format, pytest (Postgres + Redis services)
• Frontend: ESLint, TypeScript, Next.js build
• ML: pytest model tests
• E2E: Playwright (on PR)

Roadmap

See ROADMAP.md.