Video summary
Getting Hired In Post AI World (My Experience)
Main summary
Key takeaways
Business Hiring Takeaway: “Hiring Style Needs to Change” (Especially for Founding Roles)
- For a founding back-end / founding engineer (i.e., the #1–#3 hire with major decision-making responsibility), the bar is about autonomy + judgment, not just implementation ability.
- Interview design should be more personalized and less reliant on generic coding-only questions.
- Technical rounds must not be outsourced to non-technical interviewers who score asynchronously using AI. The round should be human-led (or at least human-reasoned in real time).
Playbook / Interview Format Framework (Semi-Structured)
Recommended Structure (Interviewer’s Process)
- 5–10 minutes: Introduction + background
- stack experience
- hard problems solved
- 20–25 minutes (sometimes more): Solve a company-relevant problem via whiteboarding
- interviewer watches approach and reasoning, not just the final architecture
What the Interviewer Is Evaluating
- How quickly the candidate understands requirements
- Whether the candidate can make good architecture choices and justify tradeoffs
- Awareness of AI-agent realities:
- deployment
- tooling
- model selection
- data privacy
- Ability to ask the right clarifying questions at the right time
- Reasoning clarity:
- explaining tradeoffs
- costs
- scale implications
Less “Coding Questions,” More “System Design + Tradeoff Reasoning”
The interview de-emphasizes “loaded” micro-debug questions (e.g., a dependency bug in useEffect) as the main focus.
Instead, the emphasis shifts to:
- whiteboard system design
- cost/complexity reasoning
- operational concerns:
- deployment
- caching
- rate limiting
- error handling
- scalability assumptions and how architecture changes with scale
Concrete Interview Case Study (AI + Law/Healthcare Lead Qualification Workflow)
Context
- Startup domain: AI + law (US; healthcare/law-specific lead area)
- Marketing: Facebook lead ads generate leads
- Backend receives lead data (e.g., email, phone number) and runs an AI agent workflow
Problem Statement (What to Design)
Build an AI conversational agent that:
- contacts the lead on Telegram
- runs a human-like text conversation flow
- collects specific data/files sent by the user
- stores results back in the backend
Also include an admin dashboard that:
- shows ongoing chats and what the AI is doing
- allows admins to intervene / jump into a session
- supports AI handoff (“I’m done collecting data”) and admin interrupting mid-flow
Operational / Architecture Elements Expected
- Webhooks and event-driven ingestion (Facebook → backend)
- Telegram APIs + Telegram webhooks for user replies
- System components for:
- persistence (example mentioned: Postgres)
- queuing/caching/rate limiting (example mentioned: SQS, Redis)
- AI agent orchestration (example tooling mentioned: Universal AI SDK or custom harness)
- deployment environment (example mentioned: AWS Lambda or EC2)
- dashboard support for live monitoring and admin takeover
- model choice + privacy concerns
- e.g., if a candidate mentions “open weights,” interviewer may ask about not sharing training data
Scale-Based Tradeoff Prompts (Cost/Complexity vs Usage Logic)
The interviewer explicitly tests whether the candidate adjusts architecture based on realistic volume.
Example Scale Assumptions Introduced
- < 10 leads/day
- < 20 messages/lead
- => effectively ~100–200 messages/day
Evaluation Goal
- If the candidate chooses SQS (or adds queueing complexity), the interviewer asks whether they’d change decisions given low volume.
- Intent: apply “YAGNI via scale”—justify complexity only when it’s warranted.
- e.g., removing unnecessary components
- possibly using Postgres-based queuing instead
- using Redis only if truly needed
Framework: “Dynamic Interview” Loop (Clarifying Questions + Course Correction)
- The interview is not static; the interviewer uses follow-up constraints to test adaptability.
- Candidates should:
- infer missing requirements
- ask clarifying questions early (not only after designing)
- reason through components even if they haven’t built that exact system before
Key Principle
- AI tools may “just start working” with ambiguous instructions.
- Humans must detect ambiguity and ask better questions.
AI-Era Hiring Skill Gap: “AI Is Bad at Asking the Right Questions”
Observed issue: raw AI harnesses may:
- proceed with incomplete context
- drift into a different direction, requiring later correction
The interviewer looks for a candidate who naturally:
- requests scale, constraints, and privacy expectations early
- demonstrates fundamentals that lead to better question selection
Actionable Recommendations
Candidate Preparation
- Don’t interpret “less coding questions” as “less coding/system knowledge.”
- Strong fundamentals are still required:
- programming + micro/macro system understanding
- system design competence (including database basics like Postgres)
- performance/cost reasoning and component boundaries
- Practice via hands-on learning (not only passive watching).
Interview Execution (Time + Reasoning Traces)
The interviewer also provides practical feedback observed from candidates:
- Respect the time box (example: 30 minutes); don’t overshoot with too much non-technical talk.
- Show reasoning aloud (“reasoning traces”):
- approach
- alternatives considered
- tradeoffs/cost impacts
- when a missing requirement becomes clear
This is framed as more natural for humans than for LLMs (which can’t interrupt reasoning mid-stream as flexibly).
Metrics / KPIs Explicitly Mentioned
The only explicit KPIs were system-scale quantities used for design reasoning:
- leads: < 10/day
- messages: < 20/lead
- total messages/day: ~100–200/day
No business KPIs (e.g., revenue, CAC, LTV, churn) were provided in the subtitles.
Presenters / Sources
- Presenter/Source: the interviewer/content author narrating their “my experience” (no additional named sources or co-presenters mentioned).
- Sponsor mentioned: Filmora (AI video workflow sponsor; not directly tied to the hiring/business process content).