What's shipped, what's coming.

Right now progress saves to localStorage. Adding optional sign-in lets your XP, streak, completed lessons, and notes follow you across devices. Magic-link auth is wired in /login — what's left is the sync layer that runs after auth lands.

Today Mira gets the lesson summary. Next she gets the current beat plus your recent attempts so she can answer 'why did this prompt fail my last try?' with real specificity. Built on Anthropic prompt caching to keep cost flat.

Indicative Pro tier (~€15/mo) detailed on /pricing. Anyone signed up during the launch period keeps a permanent free tier covering the current catalog. Stripe webhook + customer portal scaffolding is already in app/api/stripe.

Per-seat plan from 5 seats with a manager view showing anonymized team progress (no individual surveillance). Custom internal lessons for company-specific APIs/stack at the top tier.

PWA covers most of the use case today; native apps unlock background notifications for streaks and offline pre-download of lesson media.

One-click feedback on any lesson beat with structured tagging (factually wrong, outdated, unclear, broken code). Public-facing edit history on /changelog so contributors get credit.

An engineering-flavored PM track. Eval-first specs with the 7-section template, pricing patterns with cost-per-call math, roadmapping against a moving capability ceiling, user research that doesn't lie (Wizard-of-Oz, diary studies, refusal-acceptability testing), the PM/eng/designer working agreement, and a capstone walking one realistic feature from idea through quarterly review.

Deeper than RAG-101: why dense retrieval works at all, embedding-model selection on real data, ANN internals (HNSW M/ef tuning, IVF + PQ trade-offs), vector-DB selection by operational fit (pgvector, Atlas, Weaviate, Qdrant, Milvus, Pinecone), chunking that respects boundaries and carries metadata, hybrid retrieval + reranking with RRF code, eval discipline (precision@k, recall@k, faithfulness, per-case diffs), and a capstone walking a multilingual help-center search end-to-end.

The four LLM-specific signals (refusal-rate drift, response-length anomalies, retrieval-precision decay, tool-call distribution shifts), the trace schema for incident triage, hourly probe sets against production, the on-call playbook for AI features, observability-stack selection (Phoenix, Helicone, Braintrust, Langfuse, OTEL+Datadog), privacy in traces, and a capstone that wires it all into one real feature — turning an 18-hour incident detection cycle into 14 minutes.

The five 2026 interview formats with the patterns that score: prompt-design exercise scoring rubric, the full eval-design answer, RAG and agent whiteboard structures with seniority-signaling moves, RAG-vs-fine-tuning trade-offs, broken-prompt debugging triage, take-home expectations, and a 7-line self-assessment rubric.

SRE-grade discipline applied to LLM workloads. Per-feature attribution, token budgets, prompt caching at every stable prefix, model routing, continuous cost regression in CI, self-hosting math. Capstone walks all five habits across one real feature, dropping the bill 87.5%.

Safety as engineering practice. Direct + indirect prompt injection with the layered defense pattern, building a red-team eval set, output filtering and harm classification, jailbreak categories that persist in 2026, PII handling and data minimization, bias evaluation with fairness gates, EU AI Act tiers + enterprise procurement readiness, responsible disclosure, incident-response playbook, and a copy-pasteable pre-launch checklist.

Vision encoders + patch tokenization, document understanding (PDFs, tables), chart and screenshot Q&A, ASR/TTS production decisions, video frame-sampling + RAG, and cost-aware multimodal routing. The four production-ready multimodal workloads in 2026 with the patterns that scale.

Programmatic SEO and conversion-path improvements. Every lesson searchable in one place; new visitors get a 7-day plan; six 90-day skill profiles cover RAG, Prompt Engineering, LLM Evaluation, Agent Orchestration, Fine-tuning, Prompt Injection Defense.

11-row feature comparison with pick-by-situation guidance that concedes where each alternative wins.

AI-themed visual layer using the platform's vocabulary: animated neural mesh canvas, token-streaming heading, viewport-triggered count-ups, drawn synapse dividers, hover constellations on course cards. All motion gated behind prefers-reduced-motion.

Renamed from "AI Training Course". Header gets a Courses entry; /courses 308-redirects to /paths catalog. Mobile header gains a Search button.

AI Agents path on ReAct, tool use, planning, MCP, memory, multi-agent, safety. New /news, /resources (~30 picks), /tools (45+ entries), /compare/models. llms.txt for AI/LLM crawlers; LearningResource JSON-LD on every lesson; sitemap covering all of it.

Platform v1 with full Prompt Engineering content, code-graded prompt rubrics, in-browser Pyodide for Python lessons, dashboard with XP and streak, CertQuests handoff.