Skilluv AI Service

Internal AI microservice powering Skilluv's challenge generation, plagiarism detection, talent matching, and media processing.

🇬🇧 English (this page) · 🇫🇷 Version française

---

What is Skilluv?

Skilluv is a community platform training the African OSS generation through real contributions to real open source projects. Every completed challenge produces a verifiable artifact exportable to recruiters. Talents never pay for access — companies do.

Full product vision in the backend repository.

What this repo contains

The AI microservice of Skilluv. Written in Python 3.12, packaged with uv. It communicates with the Rust backend via gRPC (synchronous calls) and Redis Queue (asynchronous jobs). This service exposes no public HTTP routes.

Responsibilities:

• Challenge generation — parameterized challenge synthesis via LLM (Claude API)
• Plagiarism detection — code similarity detection using tree-sitter and embeddings
• Talent matching — filtering and scoring for recruiter queries
• Media processing — replay ingestion, clip generation via ffmpeg and MinIO

MVP status (v2 gRPC contract)

The v2 gRPC contract (skilluv.ai.v2 — see proto/skilluv_ai.proto) is MVP-ready. Phases M1 → M5 done, M6 (deployment) is the last step.

Phase	Scope	Status
M1	Proto unification (4 services, versioning rules)	✅ done
M2	gRPC wrappers CodeReview + Plagiarism	✅ done
M3	TalentDetection (AnalyzePerformance + SuggestCareerPath)	✅ done
M4	GenerateChallenge + GenerateVariant v2	✅ done
M5	Metrics, Grafana, alerts, e2e tests, docs	✅ done
M6	Docker image + Hetzner deploy + secrets	✅ done

Key entry points :

• Plan & decisions : docs/MVP.md (§0 = decisions arrêtées)
• API contract : docs/API-CONTRACT.md (grpcurl examples for the 4 services)
• Deployment : docs/DEPLOYMENT-HETZNER.md + docker-compose.prod.yml
• Dashboards : infra/grafana/skilluv-ai-grpc-dashboard.json
• Alerts : infra/prometheus/alerts.yml

Test suite : ~230 tests, gRPC round-trips against a local aio.server, Claude mocked (no network cost).

Companion repositories

• skilluv-backend — Rust + Axum API (the caller for gRPC and Redis Queue)
• skilluv-frontend — SvelteKit web app for talents
• skilluv-admin — SvelteKit admin panel

Architecture

                          gRPC :50051
Rust Backend ────────────────────────────►  Challenge Generator (Claude API)

                     Redis Queue
Rust Backend ──push──►  skilluv:queue:plagiarism  ──►  Plagiarism Detector (tree-sitter + embeddings)
             ──push──►  skilluv:queue:matching    ──►  Talent Matcher (filtering + scoring)
             ──push──►  skilluv:queue:media       ──►  Media Processor (ffmpeg + MinIO)
                                                            │
                                                       skilluv:result:{job_id}
                                                            │
                                                       skilluv:notifications ──pub/sub──► Rust Backend

Services

Service	Role	Communication	External dependencies
Challenge Generator	Generates parameterized challenges via LLM	gRPC sync	Claude API (Anthropic)
Plagiarism Detector	Detects similarity between code submissions	Redis Queue async	None
Talent Matcher	Filters and scores talents for recruiter queries	Redis Queue async	None
Media Processor	Ingests replays, generates clips	Redis Queue async	ffmpeg, MinIO

Quick start

Prerequisites: Python 3.12+, uv, Docker (for Redis, MinIO, and the Rust backend).

git clone https://github.com/skilluv/skilluv-ia.git
cd skilluv-ia
cp .env.example .env
# edit .env — CLAUDE_API_KEY, REDIS_URL, MINIO_ENDPOINT, etc.

uv sync
uv run python -m skilluv_ai

The gRPC server listens on port 50051.

For the full deployment guide (Docker, Hetzner, monitoring via Grafana Loki + Prometheus), see README.fr.md.

Contributing

Contributors welcome — Python engineers, ML/LLM practitioners, prompt engineers, security researchers interested in LLM prompt injection defense. See CONTRIBUTING.md and CODE_OF_CONDUCT.md.

AI-assisted contributions are welcome and encouraged (and this service literally is one — full loop transparency).

Security

For security disclosures, see SECURITY.md. Because this service handles LLM prompts and executes with API keys, prompt injection and secret exfiltration reports are especially welcome.

Monitoring

• Prometheus-compatible /metrics endpoint
• Metrics: skilluv_ai_jobs_total, skilluv_ai_jobs_duration_seconds, skilluv_ai_jobs_in_progress, skilluv_ai_external_errors_total
• Structured JSON logs to stdout, collected via Grafana Loki

License

Distributed under the GNU Affero General Public License v3.0 (AGPL-3.0). The AGPL was chosen to prevent closed forks of the AI pipeline while remaining compatible with the rest of the Skilluv stack.

Origin

Skilluv is built solo by Jeremie Zitti, a Beninese engineer. Public launch: January 2027.