Interview AI · Internal Demo

I just ran a technical interview.
This is how I did it.

A live walk-through of AI-assisted candidate interviews — from CV to scorecard, in one session.

The before

Interviewing well takes hours.

And the output depends on how fresh you are, how much you remembered to write down, and how well you compared candidates afterwards.

The old way

· Read the CV the night before, skim the JD
· Improvise questions on the spot
· Take fragmented notes while trying to listen
· Write evaluation hours or days later, from memory
· Inconsistent rubric across candidates

With AI

› Structured pre-read mapped to the scorecard
› Questions pre-generated with red flags to watch for
› Live note-taking — AI listens with you, flags inconsistencies
› Evaluation written before you close your laptop
› Same rubric, every candidate, every time

The shift

One command. Three acts.

A single entry point. The AI reads the JD, the scorecard, and the CV, then guides you through a structured process.

$ /interview senior-platform-engineer

ACT I

Prep

CV read, fit analysis, scorecard-mapped question set with red flags.

≈ 5 minutes

ACT II

Interview

Live note capture. AI listens silently, flags inconsistencies and red flags in the moment.

During the call

ACT III

Evaluate

Full scorecard evaluation, decision recommendation, auto-filed in the right folder.

≈ 2 minutes

Act I · Prep

ACT I

I.

Prep, not guesswork.

Before I join the call, the AI has already mapped the candidate to the scorecard and generated tailored questions — including the specific red flags to listen for.

01
Reads the JD + scorecard Focus attributes grouped as soft and hard technical, including any hard filters.
02
Parses the candidate CV Career arc, stack, leadership scope, education, certifications.
03
Scores fit attribute-by-attribute Strong / partial / gap — with evidence from the CV for every call.
04
Generates questions and red flags At least one question per attribute, escalating depth for hard filters.

Act I · output

Every attribute, mapped. Every gap, flagged.

A real pre-read (anonymised) — instantly tells me where to spend my 45 minutes.

Candidate A — Senior Platform Engineer · fit matrixGenerated in 3m 12s

#	Attribute	Fit	Note
2	10+ years experience	Borderline	CV says 9+, timeline ~10.5
4	Content platform expertise	Gap	No relevant experience on CV
6	Team leadership	Partial	"Lead" title; squad size unknown
8	Frontend framework at scale (HARD)	Partial	Modern stack yes; traffic unclear
9	Multi-tenant architecture	Gap	Critical for this role
12	High-traffic platform scale	Gap	Only internal/auth tooling
16	Typed API design	Strong	Type-safe stack in production

Act II · The interview

A silent partner
in the room.

I type shorthand notes between my questions. The AI acknowledges, stays out of the way, and listens for patterns.

Me

leads a team of 4 FE, 4 BE and 1 QA

AI

Noted. Team of 9 direct reports — solid squad size for a lead role. Strengthens #6 (team leadership) and #10 (squad/tech lead experience). Keep going.

Me

uses redis for better performance

AI

Noted. Aligns with CV (Redis-backed async processing). Keep going.

Act II · catch

The AI catches what I miss.

An inconsistency surfaced mid-interview — something I might not have noticed in the moment.

12:04

Me: "works with tRPC and Zod across the stack"

12:17

Me: "haven't heard of SOAP or RPC"

12:17

        Red flag
        AI: Red flag on typed API design: unfamiliar with SOAP and RPC — these are foundational API paradigms. CV lists tRPC which is literally "TypeScript Remote Procedure Call." Probe whether the tRPC usage is shallow.
      

Act II · multimodal

Whiteboard? Just paste the photo.

The AI sees what the candidate sketched — and folds it into the notes alongside everything else.

whiteboard.png

Architecture sketch (anonymised) — shown during the whiteboard exercise

AI added to notes

Frontend rendering layer
Middleware / API layer
Content layer as the backbone
CDN caching layer
Three tenants sharing components
Flag: no caching depth, no edge strategy, no content inheritance model

Act III · Evaluate

ACT III

III.

Two words: "end interview"

The AI compiles everything — notes, inconsistencies, red flags, whiteboard content — into a full scorecard-based evaluation with a decision recommendation.

01
Rates each of the 17 attributes Star / Thumbs up / Neutral / Thumbs down / No — with evidence.
02
Writes strengths and critical concerns Prioritised: HIGH / MEDIUM / LOW — so the hiring panel knows what matters.
03
Recommends ADVANCE / HOLD / REJECT With rationale and suggested next step.
04
Files it in the right folder Moves CV + evaluation into Interview/Completed/<Candidate>/

Act III · output

The evaluation writes itself.

Real anonymised output from the session — decision, rationale, and scorecard distribution.

Overall recommendation · Candidate A

REJECT

Solid Lead Software Engineer with genuine full-stack experience, strong security awareness, and honest self-assessment. Lacks the architectural depth, platform expertise, and scale exposure this senior role demands.

0 Star

3 Thumbs up

5 Neutral

5 Thumbs down

4 No

Top strengths

Honest self-assessment
Leads a 9-person squad, reports to CTO
Strong security mindset (MSc, Sec+)

Critical concerns

No platform customisation depth
No multi-tenant architecture experience
No high-traffic public-facing scale

What changed

Less time. Better evidence. Same rubric every time.

Quick comparison from this one session vs. how I'd have run it a year ago.

3h›5m

Pre-interview prep

CV read, fit matrix, question set — all in one run

2h›2m

Post-interview write-up

Full evaluation auto-compiled from live notes

17 / 17

Scorecard attributes covered

Every attribute rated, with evidence — nothing dropped

Try it yourself

Four steps. Fifteen minutes of setup. Then you're running.

The whole workflow is driven by a folder convention. No new tools to learn.

1

Create the role folder Hiring/<Role Title>/ with Job Description/, Interview/, and scorecard.md

2

Drop the CV into Interview/ PDF or DOCX. The AI reads it when the session starts.

3

Run the command /interview <role> — tell the AI your role (hiring manager, recruiter, technical).

4

Take the interview. Type shorthand notes. Say "end interview" when done. The evaluation and filing happen automatically.

Over to you

Your move.

Pick your next interview. Try it once. If it saves you an hour and catches one thing you'd have missed — it's already worth it.

Questions? Let's open the floor.

I just ran a technical interview.This is how I did it.