Question 1

What does AI-first engineering mean in your daily work?

Accepted Answer

It means I design the workflow before I pick the model. I map human steps, decide where judgment is required, then assign Cursor for code, MCP servers for live APIs, n8n for triggers, and Python for batch jobs. The LLM is one station on an assembly line I can test, log, and swap — not the whole factory.

Question 2

How is AI-first different from just using ChatGPT for everything?

Accepted Answer

ChatGPT is great for thinking out loud, but production work needs boundaries. I use models inside pipelines with structured outputs, retrieval, tool permissions, and eval gates. When a workflow repeats, I automate it; when it needs taste or accountability, I keep a human in the loop.

Question 3

When do you choose automation over an agent?

Accepted Answer

If the steps are fixed and the data shape is stable, I reach for n8n or a Python script first. Agents earn their place when intent varies, context matters, or tool selection depends on the request. I have seen teams burn weeks agentifying a flow that a cron job could own.

Question 4

What is the first thing you build on a new AI project?

Accepted Answer

An eval question set and a logging skeleton. Before prompts or UI, I write twenty questions stakeholders actually ask, plus failure cases I am afraid of. Those become regression tests, FAQ seeds, and the acceptance bar for shipping.

Question 5

How do you keep AI projects from becoming endless experiments?

Accepted Answer

I time-box discovery, then ship a narrow vertical slice with metrics: latency, accuracy on the eval set, and human review rate. If we cannot measure it, we are still in demo land. Production AI systems need a definition of done like any other feature.

Question 6

What surprised you most after moving to an AI-first workflow?

Accepted Answer

How much of the work is integration, not prompting. Wiring Gmail, CRM, GitHub, and vector stores through MCP took more calendar time than tuning system messages — but it is also what made outputs trustworthy and repeatable.

Question 7

What is an MCP server and why do you rely on them?

Accepted Answer

Model Context Protocol servers expose typed tools — search inbox, fetch ticket, create draft — that agents call at runtime. I build custom MCP layers over Gmail, GitHub, Supabase, and Stripe so the model never gets raw admin keys. It is the difference between a demo chat and an operator that can act safely.

Question 8

When should I build a custom MCP server instead of direct API calls?

Accepted Answer

When multiple agents or IDEs need the same capability, or when you want audit logs and permission scopes in one place. I wrapped our CRM lookups once in MCP and every workflow — support copilot, email triage, reporting — reused the same tool definitions.

Question 9

How do you secure MCP tools in production?

Accepted Answer

Least privilege, read-before-write, and explicit schemas. I start with search and fetch tools, add mutations only with human approval or high-confidence gates, and log every invocation with user, args hash, and result status. Agents should feel boringly safe to operate.

Question 10

Can MCP replace all my backend integrations?

Accepted Answer

No — MCP is the agent-facing surface, not your domain layer. Business rules, validation, and persistence still live in services you control. MCP translates intent into calls your backend already understands.

Question 11

What is the hardest part of maintaining MCP servers?

Accepted Answer

Schema drift. When an upstream API changes, tools break silently unless versioned and tested. I treat tool definitions like public API contracts with changelog notes and smoke tests agents run in CI.

Question 12

How does Context7 MCP fit your stack?

Accepted Answer

It pulls live library docs into the IDE during implementation so I am not guessing API shapes from memory. Paired with repo-aware agents, it cuts down hallucinated imports and deprecated patterns — especially on fast-moving frameworks.

Question 13

GitHub MCP or custom scripts for delivery workflows?

Accepted Answer

GitHub MCP for agent loops inside the IDE — open PRs, fetch issues, attach summaries. GitHub Actions for deterministic deploy and scheduled jobs. I do not ask an agent to do what CI already does reliably.

Question 14

Why Cursor instead of a vanilla editor plus API?

Accepted Answer

Cursor gives repo context, inline diffs, and skill hooks in one surface. I am faster on multi-file refactors, test scaffolding, and following project conventions because the agent sees the tree I am in — not a pasted snippet.

Question 15

How do agent skills change the way you code?

Accepted Answer

Skills are reusable playbooks — GSAP patterns, MCP setup, commit conventions — invoked instead of re-explaining context every session. I keep them narrow and versioned. A stale skill is worse than no skill because it silently teaches bad habits.

Question 16

When do you switch from Cursor to Claude Code in the terminal?

Accepted Answer

When the task spans scripts, CI configs, and logs outside the IDE sweet spot, or when I want a headless agent in a remote environment. Terminal agents excel at batch refactors with explicit file lists and checkpoint commits.

Question 17

Do AI coding agents replace code review?

Accepted Answer

They replace first drafts and boilerplate, not accountability. I still review architecture, security boundaries, and edge cases. Agents propose; humans approve merges — especially on auth, payments, and data deletion paths.

Question 18

How do you stop agents from drifting from project conventions?

Accepted Answer

Project rules, skills, and small eval checks on generated diffs. I also keep a living AGENTS or README section agents read first. When drift shows up in review, I fix the skill — not just the PR.

Question 19

What is your rule of thumb for model selection across providers?

Accepted Answer

Fast model for classification and routing, stronger model for synthesis and debugging, specialized model for multimodal ingestion. I pick by step, not by brand loyalty — and log cost per successful task.

Question 20

Why LangGraph instead of a single autonomous agent loop?

Accepted Answer

Real workflows branch, retry, and pause for humans. LangGraph lets me model those paths explicitly with memory per node instead of hoping one loop improvises correctly. It is easier to debug a graph than a transcript.

Question 21

How do you add human-in-the-loop to agent graphs?

Accepted Answer

Explicit checkpoint nodes with queues. Low-confidence classifications, outbound emails, and financial actions route to review with full context attached. Approvals feed back into logs and become labeled data for the next eval cycle.

Question 22

When is LangGraph overkill?

Accepted Answer

Single-step Q&A with retrieval, or straight API glue. If there is no branch and no memory requirement, a FastAPI endpoint with one tool call is simpler to own.

Question 23

How do you debug a stuck LangGraph run?

Accepted Answer

Step logs with inputs, outputs, and tool latency. I replay from the last good node with the same state snapshot. Most stalls are tool timeouts or schema validation failures, not model creativity issues.

Question 24

How does LangGraph pair with n8n in your automations?

Accepted Answer

n8n wakes the system — webhook, schedule, CRM trigger — then hands off to a Python worker running the graph. n8n owns SaaS glue; LangGraph owns reasoning and tool orchestration.

Question 25

Why n8n for workflow automation alongside code?

Accepted Answer

Because not every stakeholder speaks Python. n8n makes triggers and handoffs visible, versioned, and swappable without redeploying services. I use it for alerts, routing, and waking agents — not for core inference.

Question 26

n8n or Make for SaaS integrations?

Accepted Answer

Whichever matches the customer stack and ops comfort. I pick one per org and standardize. The agent layer stays the same; the trigger fabric is negotiable.

Question 27

How do you avoid silent failures in no-code automation?

Accepted Answer

Dead-letter queues, Slack alerts on error branches, and weekly replay of failed runs. Every n8n workflow gets an owner and a runbook link — same discipline as production services.

Question 28

What belongs in Python instead of n8n?

Accepted Answer

Batch inference, heavy ETL, eval runners, custom retry logic, and anything needing unit tests in CI. Python is my default for deterministic throughput; n8n is my default for visible ops glue.

Question 29

How do you design a RAG pipeline that stakeholders trust?

Accepted Answer

Citations on every claim, metadata filters, and reranking before generation. I show source titles in the UI and track retrieval misses in analytics. Trust comes from traceability, not from a confident tone.

Question 30

When does RAG fail in production?

Accepted Answer

Stale indexes, chunk boundaries that split tables, and missing ACL filters. I refresh embeddings on content changes, test retrieval with held-out questions, and enforce tenant filters at query time — not in prompt pleading.

Question 31

Do you always need a vector database?

Accepted Answer

No. Small corpora with strong metadata sometimes do better with hybrid search and keyword filters. I size the retrieval layer to the corpus and compliance needs, not to hype.

Question 32

How do rerankers change retrieval quality?

Accepted Answer

They reorder top-k chunks before the LLM sees them, which matters when embeddings alone surface near-duplicates or outdated docs. I measure nDCG on my eval set with and without reranking — the uplift pays for itself on support copilots.

Question 33

How do you ground support copilots in tickets and FAQs?

Accepted Answer

Hybrid index over help articles, resolved tickets, and internal runbooks with structured fields for product area and version. The copilot must cite doc sections and ticket IDs so agents can verify before sending.

Question 34

What is your eval strategy before shipping agent changes?

Accepted Answer

Golden questions from real users, tool-call assertions, JSON schema validation, and retrieval hit rates. I run them in CI on prompt or tool changes and block merges on regressions beyond an agreed threshold.

Question 35

How many eval cases are enough to start?

Accepted Answer

Twenty honest questions beat two hundred synthetic ones. I seed from support logs, sales calls, and the FAQ I am writing anyway, then grow the set every time production surprises us.

Question 36

How do you eval generative answers that are not exact match?

Accepted Answer

LLM-as-judge with rubrics, human spot checks, and structural checks — required fields present, citations included, banned phrases absent. Perfect text match is the wrong bar for conversational outputs.

Question 37

What metrics do you watch after launch?

Accepted Answer

Human review rate, deflection where applicable, p95 latency per step, retrieval miss rate, and cost per successful task. Sentiment alone is noise without operational counterparts.

Question 38

How do structured outputs reduce production incidents?

Accepted Answer

Downstream code stops parsing prose. Categories, confidence scores, and citation arrays are validated before any email sends or ticket updates. Failures become schema errors you can alert on.

Question 39

What goes into your agent observability stack?

Accepted Answer

Structured logs per step, trace IDs across tools, payload redaction, and replay fixtures. I want to reconstruct why an agent chose a tool three Tuesdays ago without guessing from a chat transcript.

Question 40

How do you handle PII in agent logs?

Accepted Answer

Redact at ingestion, hash identifiers where correlation is enough, and separate full-fidelity debug captures behind break-glass access. Agents touch sensitive data — logs should not leak it by default.

Question 41

Docker in an AI-first stack — necessary or optional?

Accepted Answer

Strongly recommended for workers and MCP services so dev and prod match. Containerized agents and eval runners make rollbacks and capacity changes predictable.

Question 42

How do you deploy agent APIs safely?

Accepted Answer

FastAPI behind auth, rate limits, idempotent tool design, and feature flags for model or prompt changes. I can roll back behavior without redeploying the whole front end.

Question 43

What is your incident response when an agent misfires?

Accepted Answer

Kill switch on outbound actions, replay the trace, patch tool scopes or prompts, add a regression eval, then postmortem publicly with stakeholders. Speed matters, but so does updating the eval set so it does not happen twice.

Question 44

How does FAQ content help SEO for AI engineering pages?

Accepted Answer

Search engines and users scan for direct answers. A deep FAQ block targets long-tail queries — MCP setup, RAG grounding, eval strategy — with canonical Q&A pairs that match FAQPage schema and internal linking to tools and projects.

Question 45

What is AEO and how do you optimize for it?

Accepted Answer

Answer Engine Optimization means structuring content so AI answer engines can cite you: clear questions, concise authoritative answers, schema markup, and speakable summaries. I write the way I would brief a smart colleague, not the way I would keyword-stuff a landing page.

Question 46

How does GEO differ from traditional SEO in your content?

Accepted Answer

Generative Engine Optimization focuses on being referenced in AI-generated answers — entity clarity, citations, freshness signals, and FAQ depth. I link expertise pages, show real stack choices, and keep answers updated when tools change.

Question 47

Which keywords matter most on an AI-first portfolio page?

Accepted Answer

Phrases people actually ask: AI-first engineering, MCP servers, RAG pipeline, LangGraph orchestration, n8n automation, agent skills, structured outputs, production evals, and workflow automation. I weave them into answers naturally, not as tags alone.

Question 48

Should FAQ answers sound corporate or personal?

Accepted Answer

Personal and precise wins. I share how I decide, what failed before, and what I measure — that is what humans trust and what answer engines quote. Definitions without context rarely rank or get cited.

Question 49

How long should FAQ answers be for featured snippets and LLM citations?

Accepted Answer

Roughly sixty to one hundred twenty words: enough for a story and a concrete detail, short enough to scan. I lead with the direct answer, then add one example from production.

Question 50

Do you duplicate FAQs across project and AI-first pages?

Accepted Answer

I overlap themes, not text. Project FAQs stay case-specific; AI-first FAQs explain philosophy and stack choices across work. Cross-linking beats copy-paste for SEO and user trust.

Question 51

How do agent skills differ from long system prompts?

Accepted Answer

Skills are modular, discoverable, and testable. A prompt monolith rots silently; a skill can be reviewed, versioned, and removed when a library updates. I treat skills like functions with docs.

Question 52

When do you write a custom Python pipeline instead of another agent?

Accepted Answer

When throughput, determinism, or cost caps dominate. Batch embedding jobs, nightly eval reports, and invoice reconciliation are pipelines first. Agents sit on top when judgment is required.

Question 53

How do FastAPI workers fit your architecture?

Accepted Answer

They host agent endpoints, webhooks, and tool backends with clear OpenAPI contracts. n8n and front ends call the same API humans can curl in an incident — no magic.

Question 54

What is your approach to email and inbox automation?

Accepted Answer

Classify intent cheaply, fetch context through MCP, draft with templates grounded in RAG, route low confidence to review. I never auto-send on the first week of a new flow — I measure draft acceptance first.

Question 55

How do you automate research without losing citations?

Accepted Answer

Multi-step agents: retrieve, summarize with sources, compare, critique. Each step writes structured notes with chunk IDs so the final brief is auditable — not a wall of confident prose.

Question 56

What makes a support copilot actually useful?

Accepted Answer

Accurate retrieval over FAQs and tickets, structured replies with confidence, seamless escalation, and a feedback button that feeds evals. Vanity chat widgets without grounding create more work than they save.

Question 57

How do you choose between Claude, GPT, and Gemini?

Accepted Answer

By step requirements: context length, multimodal needs, tool-calling reliability, and cost at our traffic. I benchmark on our eval set, not on public leaderboards alone.

Question 58

What is the biggest mistake teams make adopting AI-first workflows?

Accepted Answer

Shipping demos without evals, logging, or rollback paths. The second biggest is forcing one tool to do everything — monolithic agents, monolithic prompts, monolithic blame when it breaks.

Question 59

How can I hire you to map AI-first workflows to my stack?

Accepted Answer

Start from a painful repeatable workflow — support, research, ops, or delivery. I audit steps, pick the right mix of MCP, RAG, orchestration, and automation, then ship a measured vertical slice with FAQs and evals you can own. Reach out via the contact page with your stack and constraint.