Interview Question Index.md

title	Interview Question Index
created	2026-05-18
updated	2026-05-18
tags	interview-prep index cross-reference
audience	anyone doing a pre-interview review pass over the curriculum
stack	vault-internal lookup table

Interview Question Index

Pre-interview review entry point. Map common interview-question patterns to the chapter + measured anchor that answers them. Open this file 24-48 hours before an interview; follow the wikilinks to refresh on the specific chapter sections.

Format: each row carries a question pattern (paraphrased typical interview ask), the chapter that covers it, and a measured anchor — the specific number / output / commit ID that turns a soundbite from claim to evidence.

---

Tier 1 — Foundations (LLM Internals, Tokenizer, Attention)

Question pattern	Chapter	Measured anchor
"What does the model see when you type 'hello'?"	[[Week 0.5 - LLM Internals Speedrun]] Concept 1	5-stage pipeline: tokenize → embed → positional → blocks → sample
"Walk me through self-attention. Why three matrices?"	[[Week 0.5 - LLM Internals Speedrun]] Concept 2 + Phase 2	4-token by-hand worked example with row-0 softmax weights $(0.39, 0.14, 0.235, 0.235)$
"Is the same token's vector identical at different positions?"	[[Week 0.5 - LLM Internals Speedrun]] Concept 3 + Phase 3	Cosine sim on real MLX model: ~0.85-0.95, never 1.0
"Why isn't temperature=0 truly deterministic?"	[[Week 0.5 - LLM Internals Speedrun]] Concept 5	Softmax tail numerical noise; near-tied logits flip on argmax
"What's the difference between top_p, top_k, temperature?"	[[Week 0.5 - LLM Internals Speedrun]] Concept 5	Orthogonal knobs; tune in that order

Tier 2 — RAG (Retrieval + Eval)

Question pattern	Chapter	Measured anchor
"Walk me through your RAG pipeline."	[[Week 1 - Vector Retrieval Baseline]] + [[Week 2 - Rerank and Context Compression]] + [[Week 3 - RAG Evaluation]]	hybrid recall@10 = 0.998 vs dense 0.993
"When does GraphRAG win over vector RAG?"	[[Week 2.5 - GraphRAG]]	v12.4m: ALL judge 0.96, 32/0/0 W/L/T vs VectorRAG
"How do you evaluate a RAG system?"	[[Week 3 - RAG Evaluation]]	RAGAS faithfulness + answer relevance + Phoenix
"Multi-dimensional query rewriting + user-in-loop"	[[Week 3.7 - Agentic RAG]]	Self-RAG + CRAG decomposition; user-in-loop in roadmap
"Recall pipeline"	[[Week 1 - Vector Retrieval Baseline]] + [[Week 2.7 - Structure-Aware RAG]]	16/16 GT-judge on PageIndex vs Vector 0.5 / Graph 0.375

Tier 3 — Memory (Cross-Session, Multi-Agent, Two-Tier)

Question pattern	Chapter	Measured anchor
"How would you give an agent long-term memory?"	[[Week 3.5 - Cross-Session Memory]]	15/15 recall benchmark; mem0 cross-check 10/14
"How would you architect memory for a multi-agent system?"	[[Week 3.5.8 - Two-Tier Memory Architecture]]	Two-tier 85% on 15-Q benchmark vs single-tier baselines
"How do you handle long-term agent memory under contradiction?"	[[Week 3.5.8 - Two-Tier Memory Architecture]] §9.6	6-action classifier (add/update/supersede/coexist/delete/no-op); 5/5 tests PASS in 76.5s
"What did you learn building a consolidation pipeline?"	[[Week 3.5.8 - Two-Tier Memory Architecture]] §3	Idempotency / ordering / failure isolation; facts_deduplicated=2 on duplicate-scroll test
"Atomic-claim race in multi-agent systems"	[[Week 3.5.5 - Multi-Agent Shared Memory]]	guild SQLite UPDATE WHERE owner IS NULL primitive

Tier 4 — Agents (ReAct, Routing, Tools, Skills)

Question pattern	Chapter	Measured anchor
"Build a ReAct loop from scratch"	[[Week 4 - ReAct From Scratch]]	~150 LOC no-framework
"MLX Studio gateway + role map"	[[Week 4 - ReAct From Scratch]] §1.5	7-role ROLE_MAP; gateway adds 0% overhead; JANG tool=1.00 vs heretic tool=0.00
"How do you route between model tiers?"	[[Week 4.5 - Model Routing and Effort Tiering]]	3 tiers × 3 modes classifier + BART vote + 4-way bench
"How do you parallelize intent recognition?"	[[Week 4.5 - Model Routing and Effort Tiering]] §Phase 4	asyncio.gather vote layer + SQLite disagreement log + safety-bias tie-break
"How do you control inference cost?"	[[Week 4.5 - Model Routing and Effort Tiering]] §Phase 5	$-cost rate card; 4-way Pareto bench
"Single-agent vs multi-agent / sub-agent design"	[[Week 3.5.5 - Multi-Agent Shared Memory]] + [[Week 5 - Pattern Zoo]]	guild as operational tier for sub-agent handoff
"Skills implementation"	[[Week 6.7 - Authoring Agent Skills]]	Anthropic Claude Skills pattern; markdown + code package

Tier 5 — Prompt Engineering + Context

Question pattern	Chapter	Measured anchor
"How are prompt templates constructed?"	[[Week 6.85 - Prompt Template Engineering Patterns]]	5-axis design space; schema-enforce-with-retry-once
"How do you handle malformed LLM output?"	[[Week 6.85 - Prompt Template Engineering Patterns]] §2	Strip-fence + Pydantic validate + retry-once-with-error; turned ~30% fence-wrap failures to zero
"When do you use few-shot vs zero-shot?"	[[Week 6.85 - Prompt Template Engineering Patterns]] Axis 4	3 rules: zero / one / many; embedded-example for tight schemas
"Tell me about context engineering"	[[Week 6.9 - Context Engineering and Todo Mechanisms]]	4 shapes: rolling / summarized / RAG-recalled / structured-state
"Why do todo lists make models more focused?"	[[Week 6.9 - Context Engineering and Todo Mechanisms]]	Cognitive narrowing; ~300 tokens for 5-item list ~3% overhead
"Walk me through context engineering for a long-running agent"	[[Week 6.9 - Context Engineering and Todo Mechanisms]]	80-LOC TodoList primitive + 4-state semantics + budget formula

Tier 6 — Code Agents

Question pattern	Chapter	Measured anchor
"How does your code agent parse code?"	[[Week 7.8 - Code-Agent Patterns AST Coverage Mocks]]	tree-sitter AST + multilspy LSP; LSP queries ~10-50ms vs LLM ~1-5s
"Branch coverage statistics + AST instrumentation"	[[Week 7.8 - Code-Agent Patterns AST Coverage Mocks]] Concept 3	coverage.py --branch; sys.settrace mechanism
"Which code can't your agent generate unit tests for?"	[[Week 7.8 - Code-Agent Patterns AST Coverage Mocks]] Concept 4	4 classes: I/O / decorated / dynamic-dispatch / concurrency
"How do you mock?"	[[Week 7.8 - Code-Agent Patterns AST Coverage Mocks]] Concept 5	DI > patch > monkey-patch; spec-validated MagicMock
"How do you improve coverage iteratively?"	[[Week 7.8 - Code-Agent Patterns AST Coverage Mocks]] §3	Loop: missed branches → LLM prompt for edge tests; hit 92% from 65% baseline

Tier 7 — Inference Optimization + Production Infra

Question pattern	Chapter	Measured anchor
"When do you quantize a model?"	[[Week 7.7 - Quantization and Inference Optimization]]	Bit-width hierarchy; 70B at MXFP4 = 35GB fits where FP16 doesn't
"Walk me through inference optimization for real-time agents"	[[Week 7.7 - Quantization and Inference Optimization]]	3 levers: quant + KV-cache cap + memory-bound regime; M5 Pro ~30 tok/s on 20B-MXFP4
"Difference between PTQ and QAT?"	[[Week 7.7 - Quantization and Inference Optimization]] Concept 2	PTQ cheap +0.5-3% PPL; QAT for <4-bit; AWQ/GPTQ for calibration-aware
"Memory-bound vs compute-bound"	[[Week 7.7 - Quantization and Inference Optimization]] Concept 4	$\text{tok/s} = \text{Mem BW} / (N_{\text{params}} \cdot b_w)$; AR decoding is memory-bound at batch=1
"How would you observe a production agent?"	[[Week 11.6 - Production Tracing and Cost Telemetry]]	OTel + Langfuse + DuckDB; 5-10% per-call overhead; 10x debuggability gain
"Why p99 and not mean?"	[[Week 11.6 - Production Tracing and Cost Telemetry]] Concept 4	Heavy-tail; p99=7800ms while mean=850ms
"Walk me through cost attribution"	[[Week 11.6 - Production Tracing and Cost Telemetry]] §3	$C_{\text{call}} = t_{\text{in}} \cdot p_{\text{in}} + t_{\text{out}} \cdot p_{\text{out}}$; loop=30%cost+80%calls vs finisher=30%cost+1%calls
"Production LLM infrastructure (gateway, caching, fallback)"	[[Week 7.3 - Production LLM Infrastructure]]	Akshay 6-area #2+#5

Tier 8 — Production ML (CI/CD/CT, Drift, Eval Gates)

Question pattern	Chapter	Measured anchor
"Walk me through your CI/CD/CT pipeline"	[[Week 11.8 - Continuous Training and MLOps Pipelines]]	3 tiers; PSI 3-day rule → retrain workflow → eval gate → 5% shadow → ramp
"How do you detect drift in production?"	[[Week 11.8 - Continuous Training and MLOps Pipelines]] Concept 3	PSI per feature; chi-square label dist; rolling 7-day accuracy
"What's MLOps Level 3 and why does it matter?"	[[Week 11.8 - Continuous Training and MLOps Pipelines]] Concept 1	Automated retraining triggered by signals, not by humans
"Eval-gated deployment"	[[Week 11.8 - Continuous Training and MLOps Pipelines]] §2	$\text{deploy iff cand} \geq \text{prod} - \epsilon$; PR gates on RAGAS faithfulness

Tier 9 — Generative Media

Question pattern	Chapter	Measured anchor
"Fine-tune diffusion for brand-consistent generation"	[[Week 8.7 - Generative Media and Fine-tuning]]	LoRA r=16 + DreamBooth-style; 8-15 ref images; ~30 min M5 Pro training
"How do you evaluate generative output quality?"	[[Week 8.7 - Generative Media and Fine-tuning]] Concept 5	CLIP + FID + human panel; CLIP saturates ~0.32; FID ~7-10 for SDXL on COCO
"Difference between LoRA, DreamBooth, ControlNet?"	[[Week 8.7 - Generative Media and Fine-tuning]] Concept 4	Identity (LoRA) + structure (ControlNet) + style (IP-Adapter) — orthogonal levers
"Video extension + temporal consistency"	[[Week 8.7 - Generative Media and Fine-tuning]] Concept 6	SVD I2V; AnimateDiff motion module; per-frame LoRA insufficient

Tier 10 — Security + System Design

Question pattern	Chapter	Measured anchor
"Agent security threats"	[[Week 11.5 - Agent Security]]	(see chapter — prompt injection, data exfil, etc)
"System design for an agent at scale"	[[Week 11 - System Design]]	Multi-tenancy, observability, cost, fallback
"Framework shootout"	[[Week 10 - Framework Shootout]]	LangGraph vs LlamaIndex vs OAI Agents SDK head-to-head

---

How to use this index

Pre-interview review (24-48 hours before):

1. Skim this index → flag rows where you don't immediately recall the measured anchor.
2. Open the chapter via wikilink; re-read the specific Concept / Phase referenced.
3. Drill the soundbite (each chapter has 2-3 interview-soundbite entries in §7).
4. Note any FAILURE MODE entry from the chapter's Bad-Case Journal — interviewer probes failure modes more than happy paths.

During the interview:

If you're asked a question pattern that's in this index → answer with the measured anchor first, then the mechanism. "p99 latency on JANG_4M was 12 seconds" beats "p99 is generally important." The number is the engineering signal.

If you're asked a question pattern NOT in this index → that's the index's gap; flag for post-interview update.

After the interview:

Convert any newly-encountered question pattern into an index row. The index is a living document; each interview that asks something new is signal to add a row.

---

Index maintenance

This file is in the vault root. When a new chapter is added:

1. Identify which Tier (1-10) the chapter belongs to (or add a new Tier).
2. Add rows for the chapter's interview soundbites (§7 of the chapter).
3. Include measured-anchor numbers — pull from RESULTS.md in the lab repo, NOT from vibes.
4. Update the chapter's Cross-References (§9) to back-link to this index.

If a chapter's measured numbers change (lab re-run produces new numbers), update the anchor cell here too. The whole index should agree with the chapters at any commit.