Evals, Red Teaming & 30-Day Interview Drills
Interactive technical curriculum synthesized from FAANG and Tier-1 AI startup interview rounds.
Automated Evaluations, Red Teaming & 30-Day Interview Drills
Comprehensive coverage of QA-18, QA-19, QA-35, QA-38, QA-41, and the complete Section F 30-Day Mock Drill Framework. Master LLM-as-a-judge quality gates, eBPF runtime security, and candidate interview execution.
| Metric | Formula / Definition | Pass Threshold |
|---|---|---|
| Faithfulness | Verifiable claims in output ÷ claims in source context | ≥ 0.92 (Zero hallucinations) |
| Answer Relevancy | Cosine similarity of answer vs user query | ≥ 0.88 |
| Context Recall | Ground truth facts retrieved ÷ total facts needed | ≥ 0.85 |
Enforce in GitHub Actions: Block PR merge if Faithfulness drops by > 2% against the baseline test suite.
- Untrusted Data Extractor (Zero Tools): Reads raw emails and PDFs. Outputs purely typed Pydantic JSON. Has zero access to network or file tools.
- Privileged Agent: Reads validated JSON fields only. Never sees raw external prompt strings.
- Execution Isolation: Tool executions (code scripts, bash commands) run in ephemeral AWS Firecracker microVMs with link-local IP
169.254.169.254blocked.
30-Day Mock Interview Countdown Rehearsal Framework
Week 1: Serving & Memory Math
Calculate Llama-3-70B KV cache on whiteboard cold. Explain FlashAttention online softmax and PagedAttention block tables without hesitation.
Week 2: Kubernetes & Networking
Write a KEDA queue-based ScaledObject manifest. Explain IPVS vs eBPF Cilium bypass for 5,000 services.
Week 3: Agents & Protocols
Build a LangGraph state graph with checkpointer and pre-generated idempotency tokens. Whiteboard Model Context Protocol (MCP) vs REST.
Week 4: Failure War Stories & Design
Rehearse the "2 AM Production Incident" using the STAR framework. Execute full 45-minute timed whiteboard RAG designs.