Space Coders Agency

Final project of the AI Project Manager program (Wild Code School / Iliad). As a Teacher Assistant, I observed the daily challenges of intensive digital training over several months. Field issues identified: • Trainers: 6-8h to create a quest (an Odyssey learning assignment), up to 12h for complex topics, ratio of 70% preparation / 30% support • Learners: 1/3 feel behind, 1/4 are bored with standard exercises, demand for instant feedback • Managers: late detection of students at risk of dropping out (week 8-10 out of 12), lack of objective metrics Design Thinking approach in 5 phases: 1. Empathy — Interviews with 8 trainers, 42 learners and 3 managers 2. Definition — Problem statement: "How can we automate the creation of personalized content while preserving human support?" 3. Ideation — Collaborative brainstorming on Miro with the Iliad team and trainers from the Lyon campus 4. Prototyping — MVP with 3 specialized agents tested on 5 real quests (JavaScript, React, Node.js, DevOps, databases) 5. Testing — Validation with 12 volunteer learners and 3 evaluating trainers

Development carried out fully independently over 7 months (March-October 2025) under strong constraints: the receivership of the Wild Code School ruled out any additional funding. Hybrid methodology: • Macro planning with PERT/Gantt — 6 phases: Foundations, Production, Monitoring, RL, Extension, Optimization • Agile execution with adapted Scrum — weekly sprints, explicit Definition of Done, written daily stand-up • 5 roles taken on in rotation: Project Manager, Data Engineer, Data Scientist, Architect, Frontend Developer Tools and governance: • GitHub Projects in Kanban mode (To Do → Ready → In Progress → In Review → Done) • Monthly milestones (V1.0, V2.0...) aligned with the PERT milestones • Risk-Driven Development: upfront risk identification, Probability × Impact matrix, mitigation plans Custom monitoring application developed in parallel with 6 real-time dashboards: System (CPU, RAM, uptime), AI agents (latency, errors), RL (rewards, convergence), RAG (embedding quality), LLM benchmarks, Business (adoption, ROI).

Multi-agent architecture designed for scalability and maintainability, with local LLM execution to control costs and data confidentiality. BaseAgent design pattern: All agents inherit from a common abstract class that enforces a single contract: request validation, RAG enrichment, output formatting. Each agent adds its own specifics: tailored system prompt, business tools (Airflow DAG generation, model recommendations, etc.). 9 specialized agents: Data Analyst, Data Scientist, Data Engineer, AI Engineer, DevOps, Web Designer, Cybersecurity, QA Engineer, Cloud Architect. LangGraph orchestration: 7-node workflow (validation → enrichment → model selection → generation → check → formatting → output) with automatic retry and fallback to an alternative model. Success rate >95%. Hybrid RAG pipeline: Dense search (all-MiniLM-L6-v2 embeddings) + sparse search (BM25) with MS MARCO reranking. Base of ~15,000 chunks. Redis cache with a TTL of 30 days. Reinforcement Learning — From Q-Learning to LinUCB: Tabular Q-Learning showed its limits (1,500 states, slow convergence, no generalization). LinUCB (contextual bandit) learns 3× faster, generalizes across similar situations, and uses 23× less memory.

Containerized infrastructure with full CI/CD, an accessible interface and regulatory compliance (GDPR, RGAA 4.1 — the French accessibility standard). Infrastructure & CI/CD: • Multi-service Docker Compose architecture with dependency orchestration • GitHub Actions pipeline: lint, tests, build, automated deployment • Built-in observability: structured logs, Prometheus metrics, alerts User interface: • React with a modular component architecture and optimized state management • Native WCAG 2.1 / RGAA 4.1 accessibility: keyboard navigation, contrast, text alternatives • Light/dark modes, internationalization (FR/EN) • Simplified UX journey: agent choice → topic → options → streaming generation (SSE) Compliance & Quality: • GDPR: record of processing activities, legal basis, data minimization, exercise of rights • Cybersecurity: JWT authentication, RBAC, SCA scans in CI, CORS/CSRF/XSS protections Rigorous A/B tests over 14 days (320 sessions) with random assignment stratified by level. Adaptive LinUCB vs static hyperparameters: significant improvement on all indicators (p<0.01).

The final result: a complete platform for generating personalized learning quests, usable by Wild Code School trainers. User journey: 1. Selection of the specialized agent (Data Engineer, DevOps, etc.) 2. Choice of technology from the catalog (50+ technologies available) 3. Configuration of options (level, duration, number of exercises) 4. Streaming generation with real-time display 5. Export and integration into Odyssey (the WCS learning platform) The interface also includes a conversational chatbot to refine requests and a feedback system for continuous improvement through LinUCB.