Vela’s AI screen delivers just three candidates per posting.

How Vela Filters Candidates

Vela's screening engine begins interviews about five days after a posting goes live, and most searches deliver a shortlist within three weeks: three candidates, each with a written case and the evidence behind every rating. Most hiring funnels widen at the top and narrow by attrition: resumes pile up, someone skims keywords, a few phone screens happen, and the rest is calendar negotiation. Vela inverts that shape. The company's engine does not wait for applications; it hunts across a talent index built from the open web, including professional profiles, portfolios, code, and writing, reaching people who are not looking and not trapped on a single platform. Three agents run that top of funnel around the clock. When a candidate signals interest, a screening agent holds a real two-way conversation, then scores them against a scorecard the hiring team has already locked. The output is not a stack of resumes. It is a short, ranked shortlist of three people maximum, each accompanied by a comparable documented case.

The funnel moves through distinct stages, each designed to filter on signal rather than volume. First comes sourcing: continuous, passive, and matched by skill and trajectory instead of keywords. Second is the screening conversation, an actual dialogue, not a form, where the agent probes for the three variables Vela's own placement data shows carry the most weight. Third is the scorecard evaluation, where every rating is tied to evidence so the hiring team can audit the judgment. Fourth is the shortlist delivery: three candidates, written cases, a clear ranking. Fifth, reference checks run on the same process Vela uses for its own hires. Sixth, client interviews begin typically within five days of a posting going live, and most searches deliver that shortlist inside three weeks. Seventh, post-placement check-ins at day 7, day 30, and day 90 catch early misalignment before it hardens into attrition.

What distinguishes the screen is not the AI but the judgment poured into the scorecard. Vela's founders have run enough searches to know where AI sourcing goes wrong, how it surfaces noise when the rubric is generic, and tune each scorecard across many searches so it surfaces the right people instead. The screening agent does not improvise; it executes a calibrated rubric that asks: what have you actually owned, and for how long (scope, not titles)? What did you do when it went wrong? How do you make decisions nobody is there to check? Can you lead a call with a US client and be understood, a working requirement, not an accent test, because it is most of the job in most of the seats they fill? Those three questions filter out candidates who look strong on paper but lack the operational judgment the role demands.

The result is a funnel that says no more often than it says yes. Agencies typically present a slate and let the client decide; Vela presents a conviction. That conviction is backed by a process that replaces the DIY grind, where founders lose up to half their week to recruiting, and the agency fee of up to a quarter of first-year salary with a flat-fee engine that sources, screens, and schedules. The companies using it — Netflix, LinkedIn, SanDisk, Zoox, Scale.AI, Via, Pony.AI, AWS — are not buying a filter. They are buying a recruiter that never sleeps, never forgets to follow up, and never lets a strong candidate slip because no one had the hours to chase them.

What Vela Is Actually Hiring For

Vela's public job listings tell a more precise story than the headline count suggests. The company's Y Combinator page states it is "hiring for 4 roles in operations and engineering," and the Work at a Startup board confirms "Vela is hiring for 4 jobs" as of the W26 batch. Yet the two roles documented in detail, AI Operations Associate and AI Recruiting Coordinator for Executive Search, reveal a hiring priority that leans heavily toward human-in-the-loop oversight of AI decision-making, not pure model development.

The AI Operations Associate role appears across multiple channels with consistent requirements. The Work at a Startup posting lists it as remote in India, night shift (6:30 PM to 6:30 AM IST), six days a week, at 6–10 LPA CTC. A LinkedIn posting from June 2026 shows the same role in Bengaluru with "Over 200 applicants." The Y Combinator job description defines the work concretely: review AI-generated scheduling actions before they reach top search firms and their C-suite candidates; make real-time judgment calls on ambiguous scenarios, such as tone for a CHRO, whether to reschedule, how to handle missing thread context; catch subtle errors that matter, such as wrong time zones, misread intent, off-tone language. The company's own description adds context: "Our AI makes thousands of context-dependent decisions every day. For example: 'John will take a 7 AM call with a CEO, but not with a peer.'" This is not data labeling. It is operational quality control for an autonomous coordination system that interfaces directly with executive-level stakeholders.

The coordinator role, listed on both Y Combinator and Work at a Startup, sits adjacent to the same problem space. Executive search firms operate on reputation and nuance; a scheduling error with a C-suite candidate carries reputational risk that a generic calendar tool cannot absorb. Vela's product — "AI that handles the tedious coordination work that drains professionals' time, starting with scheduling" — targets precisely this layer: parsing conversation threads, juggling time zones, handling calendar conflicts, and managing the subtle social dynamics of who should flex for whom. The coordinator role likely involves supervising the AI's output in live search engagements, where the cost of a misread tone or a missed cultural cue is a lost placement.

The remaining two roles in Vela's that count are not individually detailed in the public postings. Given the company's two-person founding team (Gobhanu Sasankar Korisepati and Saatvik Suryajit Korisepati, per Y Combinator) and its stated backers, including Y Combinator, CRV, Rebel Fund, and angels from OpenAI, the engineering hires are almost certainly focused on the scheduling intelligence stack: context parsing, conflict resolution logic, and the preference-learning layer that lets the system know John takes 7 AM CEO calls but not peer calls. It is hiring operators who can sit inside the loop of an autonomous agent making thousands of context-dependent calls a day — and say "no, that's wrong" when the AI misreads a time zone or a power dynamic. That is the bar.

The Qualifications That Matter

Vela has not published its rubric, its panel composition, or its pass rates. The broader frontier-AI hiring market, however, has generated enough candidate-reported data to sketch what a high bar looks like in 2026. Across 300 reports compiled by Dataford, the median process at leading labs runs four rounds and carries a 53 percent positive-experience rate. The format varies by organization: OpenAI, Anthropic, DeepMind, Meta, and xAI each test different combinations of codebase navigation, research presentations, live coding with AI tooling, and deep technical questioning. Mlmentorship.com summarizes the pattern bluntly: prepare for the format, not the logo. Candidates who treat the interview as a generic "AI knowledge" exam tend to underperform; those who rehearse the specific artifact the lab requests, such as a pull-request walkthrough, a 15-minute research talk, or a debugging session inside an unfamiliar repo, advance.

Portfolio evidence has shifted from toy demos to production-grade signals. Dev.to's 2026 hiring-manager survey and BuildAIQ's case-study framework converge on five project archetypes that consistently move the needle: a model-serving stack with observability and cost controls, a data-pipeline that handles schema drift, an evaluation harness that measures regression on held-out tasks, a fine-tuning run documented with compute budgets and ablation tables, and a safety/red-teaming report with reproducible findings. Atlia Learning's hiring-manager guide adds that each artifact must be packaged as a case study, including problem statement, constraints, decisions, and measurable outcomes, not a notebook dump. GitHub hygiene (conventional commits, CI passing, issue-linked PRs) is treated as table stakes, not a differentiator.

Vela's "selective opportunities" language aligns with the self-selection mechanics observed in Match 2 career-site data: when candidates see a three-quarters-or-better match to posted requirements, application volume rises fivefold; below 40 percent match, 19 in 20 unqualified applicants opt out before submitting. The implication for Vela applicants is straightforward: read the requisition literally, map every "must-have" to a verifiable artifact, and skip the role if the gap is structural. The same data shows that intelligent applications now carry a rationale report, the original resume, and an AI-generated match explanation; candidates who pre-write their own rationale, mapping each requirement to a shipped system, a paper, or a measurable outcome, hand the screener a decision document rather than a puzzle. Candidates who clear frontier-lab screens in 2026 demonstrate three layers: fluent production engineering (not just modeling), documented judgment on trade-offs (cost, latency, safety, maintainability), and the ability to communicate both to a mixed technical/non-technical panel. Until Vela releases its own criteria, that proxy is the only grounded target applicants can aim for.

How Applicants Are Adapting

Candidates facing Vela's multi-stage screen are deploying a playbook that has crystallized across the AI hiring landscape. The core tension: generative AI has made it easier to generate polished interview answers, but Vela's process, like those at other frontier labs, is designed to detect the difference between rehearsed output and demonstrated judgment.

Preparation now starts with the model itself. MIT Sloan research documented that candidates routinely feed role specifics, organizational context, and their own resumes into large language models to generate anticipated questions and personalized answers. In one study, candidates who used GenAI for prep received higher overall interview performance ratings than unassisted peers, but the same research warned that polished, parroted responses can mislead hiring managers into attributing knowledge the candidate doesn't actually possess. Vela's later stages, which probe operational judgment, are built to expose that gap.

The most disciplined applicants treat preparation as a structured program. Shortlistd's eight-step framework, researched across hundreds of AI interviews, allocates roughly six to eight hours: company and role research (30–60 minutes), experience inventory (60–90 minutes), STAR-method response drafting (90–120 minutes), question anticipation (60 minutes), verbal practice (60–90 minutes), platform simulation (30–60 minutes), environment optimization (30–45 minutes), and mindset work (15–20 minutes). Practice platforms like Big Interview, InterviewStream, Yoodli, and VMock have become standard. The "three levels of practice" (content, fluency, simulation) mirror how engineers benchmark models: static eval, then live inference, then adversarial stress test.

Technical hygiene has become a make-or-break factor. Shortlistd found technical issues are the number-one preventable reason candidates fail AI interviews. The checklist is specific: camera at eye level, microphone tested, wired internet preferred, no background processes eating bandwidth. Upwork's guide adds that most AI interviews cannot be paused or restarted. Candidates who treat the AI screen like a casual call, with poor lighting, phone audio, and notifications firing, signal operational sloppiness before they answer a single question.

Content strategy has shifted toward keyword density and structural rigidity. AI evaluators weight content relevance highest, then communication clarity, depth of experience, keyword matching, and response structure. They do not score appearance, charisma, cultural fit, or nervousness. Successful candidates now reference the job post explicitly in answers, address every sub-part of multi-part prompts, and constrain responses to 60–90 seconds, long enough for specificity but short enough to avoid the rambling penalty. The "keyword game" exercise, which forces every answer to include five role-specific terms, has become standard prep.

Portfolio proof is replacing credential signaling. Fonzi's analysis of AI-startup hiring shows that a portfolio of concrete projects, including real-world applications, clear impact descriptions, and minimal jargon, carries more weight than brand-name degrees. Candidates are building mini case studies: problem, constraint, approach, result, lesson. They prepare to discuss ethical trade-offs, such as fairness, bias, and consent, because frontier labs now probe those dimensions directly.

A meaningful minority is opting out. Greenhouse's April survey of 1,200 U.S. job seekers found nearly four in ten have withdrawn from a process over an AI interview. Some blacklist employers that require them. One candidate with nystagmus (involuntary eye movement) described the AI interviewer repeatedly interrupting to demand she "look straight ahead," rendering the session unusable. The ACLU has warned that predictive tools analyzing facial, audio, or physical interaction increase rejection risk for disabilities, race, and other protected characteristics. A May working paper found large-scale racial disparities when employers share hiring algorithms; Black and Asian candidates were disproportionately affected.

The adaptation curve is splitting the market. Candidates who master the AI-first screen, characterized by structured, keyword-aware, technically clean, portfolio-backed approaches, advance to human rounds where they must then demonstrate the judgment the AI couldn't assess. Those who refuse the screen remove themselves from consideration at Vela and peers. Those who game the AI prep without the underlying depth get caught in the later stages. The screen works as designed: it filters for the intersection of technical fluency and operational maturity.

Why This Process Matters

Vela's screen is not an outlier. It is the leading edge of a structural shift that has already rewired how frontier AI companies hire — and the data shows the shift is accelerating.

Entry-level hiring in tech has collapsed. Ravio's compensation database, covering 400,000 employees across 1,500 companies, recorded a 73 percent drop in P1/P2 hiring in 2025 while AI/ML hiring nearly doubled year-over-year. Stanford's Digital Economy Lab found employment for software developers aged 22 to 25 fell nearly one-fifth from its 2022 peak by September 2025, while older cohorts kept growing. SignalFire put new-graduate hiring at the largest tech firms 65 percent below 2019 levels, even as overall engineering hiring dropped only 11 percent. The graduation-to-job pathway for computer science majors is materially harder than it was in 2022 or 2023.

The roles that survive have been "seniorised." PwC's 2026 Global AI Jobs Barometer, built on more than one billion job ads, found AI-exposed entry-level postings are seven times more likely to ask for senior-level skills — stakeholder management, strategic decision-making, mentorship — than they were in 2019. Those "seniorised" roles grew 35 percent since 2019; other entry-level roles shrank 10 percent. The entry-level job, as Harvard Business School's Joseph Fuller said, may now look like a second or third job would have looked five years ago.

Degrees are losing signal. The percentage of AI-augmented jobs requiring a degree fell 7 percentage points between 2019 and 2024, from two-thirds to 59 percent. For jobs AI automates, the drop was 9 points, from 53 percent to 44 percent. Employer demand for formal credentials is declining across the board, but the decline is steepest in AI-exposed roles. For the most competitive research positions, PhD requirements remain firm, yet the number of new AI PhDs in the US and Canada rose 22 percent from 2022 to 2024, and those graduates increasingly took academic jobs, not industry ones.

Skills requirements are changing 66 percent faster in AI-exposed occupations than in the least exposed, up from 25 percent faster just a year earlier. Four of the top ten skills in AI postings are not machine learning skills at all — scalability, automation, workflow management, project management — the work of putting a model into production and keeping it there. Retrieval-augmented generation mentions grew 337 percent year-over-year. Prompt engineering grew 261 percent as a skill line inside other requisitions, not as a standalone title. AWS mentions grew 1,358 percent from the 2013–15 baseline.

The supply-demand mismatch is severe. Global AI talent demand outpaces supply three-to-one; companies posted over 1.6 million AI roles worldwide against roughly 518,000 qualified candidates. Over 90 percent of global enterprises project critical AI skills shortages in 2026, risking $5.5 trillion in lost market performance, a sum roughly equal to Japan's annual GDP. Financial services and healthcare now wait half a year to fill a single senior AI role. North America pays the highest salaries for senior AI engineers. The median AI role pays substantially more than other roles, per LinkedIn data cited by CNBC. Key salary benchmarks:

Companies are responding by retraining rather than replacing. The New York Fed's surveys found only 4 percent of service firms and zero manufacturers had laid workers off because of AI in the previous six months, while 15 percent of service firms hired fewer people because of AI and 13 percent hired more. The Fed's economists wrote that existing workers are much more likely to be retrained than replaced. More than three-quarters of employers plan to upskill staff for AI-augmented roles. AWS Machine Learning Specialty certification immediately boosts salaries 10–15 percent for cloud-ML hybrid roles without requiring a degree.

Vela's multi-stage screen — technical depth, operational judgment, demonstrated AI-augmented output — mirrors what the market now demands. The screen filters for the exact profile the data shows is scarce: candidates who can ship production systems, not just train models. As agentic AI advances and more companies redesign workflows around autonomous agents instead of bolting AI onto old processes, the premium on that profile will only widen. The firms that formalize this filter now will spend less time hiring and more time building.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.