AI Engineering labs - 17 hands-on challenges

Solve 17 real AI Engineering challenges in an isolated container in your browser. No setup, no fake shells.

• Right-Size the Local Model - beginner · Recover a local Ollama service configured for a model the host cannot safely serve.
• Repair the RAG Vector Contract - intermediate · Fix an index/query embedding dimension mismatch without corrupting vectors.
• Find the Missing MLflow Runs - intermediate · Point experiment code at the team's intended tracking backend and record complete evidence.
• Restore the Approved Evaluation Data - intermediate · Recover a DVC-tracked dataset that drifted outside the Git-DVC workflow.
• Stop the Agent Secret Leak - advanced · Remove a credential from terminal-agent logs and replace environment dumping with safe audit records.
• Clean a Churn Dataset Before Training - beginner · Fix real data-quality problems in a churn dataset before any model can trust it.
• Pick the Right Model the Right Way - beginner · Fix an evaluation setup that would ship the wrong classical model.
• Diagnose Drift with a Confusion Matrix - intermediate · Catch a model quietly failing in production due to real-world drift.
• Trace Every LLM Request Like Production Would - intermediate · Replace ad-hoc logging with real structured request tracing.
• Catch a Hallucinating RAG Pipeline - intermediate · Use automated evaluation output to find and fix an unfaithful RAG answer.
• Fix a Slow Local Model Before Blaming the Model - beginner · Diagnose and fix a latency problem that has nothing to do with model quality.
• Stop an Agent From Obeying a Document - advanced · Defend a RAG-style agent against instructions hidden in retrieved content.
• Close the Gaps in a PII Redactor - intermediate · Harden an output redactor that only catches the easy cases.
• Configure Guardrails That Actually Guard - intermediate · Fix a permissive config-driven validation policy for LLM input and output.
• Tune Continuous Batching and KV-Cache - advanced · Fix a serving configuration causing terrible simulated throughput.
• Route Around a Dead Primary Model - intermediate · Fix an LLM proxy with no working fallback path.
• Make a Service Recover Like Production Should - advanced · Fix a health check and crash-recovery setup that both lie.