Viet Nguyen spent a Friday night decoding what makes Vercel's best interviewer actually good. Three hours later he had a six-slide deck the leadership team turned into the foundation of an interviewer training program. Total cost: a ChatGPT prompt, a stack of transcripts, and a refusal to wait for the perfect solution.
Viet Nguyen (Head of Global Recruiting Ops at Vercel) joined Nolan Church and Siadhal Magos on 10x Recruiting (more episodes on the 10x Recruiting hub) for the third instalment of the "How I AI" series. The episode pulls back the curtain on two weekend projects that reshaped how Vercel hires: a reverse-engineering of their best interviewer's judgment, and an automated back-channel reference system that flags warm signal at the top of the funnel. Both are small, fast, and shippable in hours.
This recap turns Viet's mindset into a model. Curiosity-driven experiments. Progress over perfection. Anyone with a transcript stack and a willingness to prompt-engineer can do this.
Reverse-engineering an A-plus interviewer
Viet noticed something that most teams notice and then ignore. One Vercel engineer named Gaspar had a track record where 87% of his "strong yes" ratings became hires who then performed at the top of their cohort. Most teams treat this as mysterious talent. Viet treated it as decodable data.
The experiment: pull every interview transcript Gaspar had ever recorded. Match each transcript to the hiring outcome (offer extended, candidate quality 12 months in, level of seniority hired into). Feed the matched dataset into ChatGPT with the prompt "what behaviors does this interviewer see in the strong candidates versus the weak ones."
In this new world, literally anybody can cook. You just need imagination and the ability to reason through how you might do something.”
Five behaviors that predict strong candidates
The ChatGPT analysis produced five concrete behavioral patterns Gaspar unconsciously evaluated in every interview. Things like how the candidate framed the problem before solving it, how they articulated trade-offs in the solution, how they handled the moment they realized they were wrong, whether their reasoning was first-principles or pattern-matched, and how they engaged with the interviewer's pushback.
The kicker Viet emphasized: these five patterns held across coding interviews AND values interviews. The behaviors were not domain-specific; they were the underlying signal Gaspar used to evaluate intelligence and operating quality regardless of interview type. That made the framework portable to every interviewer on the team.
This is the work most recruiting teams never get done because the analysis seems too hard. The reality is the data was already there in the transcripts. The cost was three hours and one ChatGPT prompt away.
Turning it into interviewer training
Viet built the insights into a six-slide deck using V0. Slides one and two: Gaspar's track record (the 87% conversion stat, the distribution of his ratings, the cohort performance). Slides three through five: the five behavioral patterns with example transcript excerpts. Slide six: the implications for interviewer training.
The leadership team reviewed the deck. The five patterns became the foundation of Vercel's interviewer training program. New interviewers now learn what to listen for, not just how to ask questions. The hidden judgment of the A+ interviewer became transferable curriculum.
The compounding payoff is structural. As more interviewers calibrate to Gaspar's bar, the average interview quality across the team rises. That lifts hire quality without adding headcount or burning out the senior interviewers.
Automated back-channel references
The second weekend project tackled the inbound funnel. Every recruiting team has done the manual version: a senior candidate's resume shows three years at Stripe, you wonder if anyone on the current team overlapped with them, you ask around. Viet automated that question.
The mechanic: a webhook fires when a new candidate applies. An AI parses the resume's company-and-date history. The system cross-references against the Vercel employee database to find anyone who worked at the same company during an overlapping period. If a match exists, the system pings that current employee with a quick "do you know this person, any take we should hear before we move them forward."
The output is not perfect. People overlap at large companies without knowing each other. Viet's framing: treat the matches as a net for positive signal, not as a gating filter. When the match comes back warm, that candidate moves to the top of the recruiter's pile. When it comes back cold or empty, the candidate goes through the normal process.
The debrief summarizer coming next
The project Viet is shipping next addresses an old recruiting pain: debrief meetings where everyone reads their scorecards aloud and 45 minutes evaporate before any real discussion happens. The plan: pipe the scorecard data into an AI agent before the meeting starts. The agent generates the summary and identifies the themes (the strong yes's, the weak no's, the disagreements worth resolving).
The interviewers validate the summary in the first five minutes, then spend the remaining 40 on the actual decision. The summary portion was the expensive part of every debrief, and it was also the part AI can do in seconds. Viet is using Metaview's structured interview data plus an internal AI agent to make it real.
The pattern across all three projects: identify a workflow where the human time is spent on aggregation rather than judgment, then automate the aggregation so the human time goes to the judgment.
Progress over perfection as an operating model
The thread connecting Viet's projects is not technical. It is philosophical. None of the systems are perfect. The back-channel system has false positives. The interviewer-decoder analysis misses nuance. Both produce measurable improvement anyway.
Perfect is the enemy of progress. We just need to make incremental progress every day because that's better than doing nothing.”
The recruiting teams that ship 50% solutions on Friday nights compound faster than the ones running quarterly planning sessions for the elegant version. The advantage flows to teams that experiment. The cost of a failed experiment is one weekend. The cost of waiting for the perfect system is six months of foregone improvement.
Where AI gives recruiting teams use
The pattern Viet's projects model is now repeatable. Capture the interview signal as structured data, build small agents on top of it, ship weekend versions and iterate.
Metaview Notetaker captures every interview as structured transcript and scorecard data, which is the input layer for any "decode the interviewer" or "summarize the debrief" experiment. Reports surfaces patterns across the interview data so the manual prompt-engineering on a single interviewer scales to the team. Application Review handles the inbound volume so the senior team has time for the experiments. For the AI-augmented recruiter pattern in depth, see claude-for-recruiters. For the interviewer-quality angle that makes the rest of this possible, see good-interviewer-bad-interviewer.
Numbers from Metaview's 2026 AI & Hiring Alignment Report, surveying 505 recruiting leaders and hiring managers across North America and EMEA. The 68% vs 49% alignment gap is what Viet's interviewer-decoder framework produces in practice. When the system surfaces what the best interviewer is unconsciously evaluating, every other interviewer on the team starts each search with a sharper bar.
The operating shift
Three concrete moves any recruiting ops leader can run this month:
One: pick your A-plus interviewer and decode them. Pull their last 12 months of interview transcripts, match to outcomes, feed it to ChatGPT with a clear behavioral-pattern prompt. You will get five-to-seven specific patterns. Convert them into the foundation of your interviewer training program.
Two: automate the back-channel for the inbound funnel. Webhook on application, AI-parse the resume, cross-reference against employee history, ping the colleague. Even a 50% accurate version surfaces enough warm signal to move the right candidates first.
Three: ship weekend versions, not quarterly initiatives. The teams that experiment iterate faster than the teams that plan. Pick the most painful workflow, ship a Friday-night prototype, learn next week, ship the next version.
The recruiting ops leaders who operate at this cadence out-iterate the ones still waiting for the perfect solution. That is the operating shift.
Bring Metaview into your hiring stack.
Live notes, structured scorecards, and ATS sync - set up in under 10 minutes.
Frequently asked questions
How did Viet decode Vercel's best interviewer?
He pulled every interview transcript his best interviewer (Gaspar) had recorded, matched each one to the hiring outcome, and fed the matched dataset to ChatGPT with a prompt asking what behaviors the interviewer saw in strong candidates versus weak ones. Three hours on a Friday night. The output: five behavioral patterns that became the foundation of Vercel's interviewer training program.
What is Gaspar's strong-yes conversion rate, and why does it matter?
87% of candidates Gaspar rated "strong yes" got offers and performed well at Vercel. That conversion rate is meaningfully higher than the team average and is what made Gaspar's judgment worth decoding. The data in his interview transcripts was the proof that his evaluation criteria were real and transferable.
How does the automated back-channel reference system work?
When a new candidate applies, a webhook fires. An AI parses the resume's company-and-date history. The system cross-references against the Vercel employee database for anyone who worked at the same company during an overlapping period. Matching colleagues get pinged with a quick "do you know this person, any take we should hear." Warm responses bump the candidate to the top of the recruiter's pile.
What is Viet's "progress over perfection" philosophy?
Ship the 50% version this weekend, iterate next weekend, beat the team still planning the elegant version six months from now. None of Viet's systems are flawless. The back-channel system has false positives at large companies. The interviewer-decoder analysis misses nuance. Both produce measurable improvement anyway. The cost of a failed experiment is one weekend; the cost of waiting for perfect is months of foregone improvement.
What is Viet shipping next?
A debrief summarizer. The agent pipes scorecard data into an AI before the meeting starts, generates the candidate summary, and identifies the themes (strong yes's, weak no's, disagreements worth resolving). The interviewers validate in five minutes, then spend the remaining 40 on the actual decision. The summary portion was the expensive part of every debrief and the part AI can do in seconds.
