Skip to content

Portfolio Overview

SoccerPredictAI is a production-style end-to-end MLOps system for football match outcome prediction.

This section is designed for hiring managers, technical interviewers, and ML/MLOps engineers who want to evaluate the project quickly without navigating the full documentation.

What this system demonstrates

Capability Where to look
One-page summary Pitch
ML methodology & results Results
System architecture & design decisions Architecture Tour
Trade-offs & out-of-scope choices Trade-offs
Key ADR digest Decisions
AI-augmented dev workflow AI Workflow
ML System Design course coverage map ML System Design
Role Recommended path Time
Recruiter / Hiring Manager Pitch only 2 min
ML Engineer Pitch → Results → Architecture Tour 10 min
MLOps / Platform Engineer Architecture Tour → Decisions → Trade-offs 15 min
Engineering Manager Pitch → Trade-offs 5 min
Technical Interviewer See deep-dive path below 15–20 min
Engineer reviewing code Quickstart → Code Structure 20 min
AI/Agentic workflow AI Workflow → Customization Layer 10 min

2-minute path

"What is this project and is it real?"

v1.0 completed May 2026. All criteria in Requirements — Definition of Done are met. Status is the canonical source.

  1. Pitch — what the system is and what it demonstrates.
  2. Implementation Status — what is built vs planned.
  3. Key facts:
  4. Data source: WhoScored.com (Airflow scraper → PostgreSQL)
  5. ML: XGBoost classifier, temporal-split validation, MLflow tracking
  6. Serving: FastAPI + Celery async, 564 automated tests
  7. Infra: Docker, Kubernetes/Helm, GitLab CI, SOPS secrets
  8. Monitoring: Prometheus /metrics; Evidently drift detection; ML quality monitor; 2 Grafana dashboards
  9. AI workflow: GitHub Copilot customization layer with scoped instructions, skills, hooks, and audit cycles

Technical deep-dive (15–20 minutes)

"Can this person design systems, not just use frameworks?"

  1. Architecture Trade-offs — documented decisions with alternatives considered
  2. ML Problem & Baseline — task formulation, why beating the bookmaker matters
  3. Feature Engineering — leakage-safe design, offline/online parity
  4. Model Contract & Signature — input/output schema enforced at boundary
  5. CI/CD Quality Gates — what runs before code ships
  6. ADR Decisions — orchestration, data versioning, serving modes

What this project proves (by competency)

Competency Evidence
Reproducibility dvc repro from clean checkout → same model. Verified by DVC lock + MLflow run IDs.
Validation rigor Temporal split enforced in code, tested with hypothesis property tests.
Serving design FastAPI with Pydantic schemas, sync + async via Celery, health endpoints.
Deployment readiness Docker multi-stage, K8s manifests, Helm charts, GitLab CI pipeline.
Observability thinking Prometheus /metrics (9 metrics), Celery queue stats, Evidently drift detection (daily DAG), Grafana dashboards deployed, alerting runbooks documented.
AI-augmented engineering GitHub Copilot customization layer with scoped instructions, skills (audit-system, error-analysis, train-serve-skew-check), hooks, and audit cycles. AI Workflow.
Operational maturity 564 tests (unit / property / service / contract / load), SOPS + age secrets.
System thinking C4 diagrams, ADRs, explicit layer contracts, no cross-layer shortcuts.

Full documentation

Full system documentation is at docs.soccer.dmitryivanov.dev (planned: live deploy post-v1.0).