The 90-Second Video That Can Beat a Perfect AI Résumé

The Initial Filter

Ninety-nine percent of Fortune 500 companies now use applicant tracking systems that reject an estimated three-quarters of resumes before a human ever sees them, per The Bridge Chronicle data cited by the University of Chicago's Polsky Center. Over 90% of job seekers use tools like ChatGPT to craft applications, per Huntr's second-quarter 2025 report. The checklist (degree tier, publication count, GitHub stars, prior employer brand) has become noise. Companies that consistently hire well have stopped trusting it.

At those levels, the initial screen is rarely a resume scan. It's a proxy for: can this person operate at the density current teams operate at?

Three proxies dominate. First, evidence of shipping at scale: not a side project, not a research prototype, but a system that served production traffic or trained a model that shipped. Second, depth in a bottleneck discipline: distributed training, inference optimization, data quality pipelines, evaluation infrastructure. Third, a traceable collaboration signal: a merged PR in a major open-source repo, a maintainer role, or a reference from someone the hiring team already trusts.

The checklist still exists. HR systems require it. But the decision to advance a candidate increasingly hinges on whether a current engineer will vouch for the work. At Anthropic, the referral path isn't a perk — it's the primary signal that a candidate's claimed contributions survive contact with reality.

Companies raising the bar use take-homes that mirror production constraints: latency budgets, memory limits, fault tolerance. The filter becomes the work itself. Candidates who advance treat the application as a technical audition, not a paperwork exercise.

The Roles the Market Demands

Databricks' latest week added 56 roles weighted toward enterprise sales leadership (Directors and Sr. Directors for vertical-specific Lakebase motions). Anthropic's 44 new postings cluster in research engineering and distributed systems, including Staff+ RL Data Platform and Pre-training Distributed Systems Tech Lead roles. Both companies hire in cohorts that reflect product priorities: Databricks pushing vertical go-to-market, Anthropic scaling research infrastructure.

Early-stage AI firms building a technical moat tend to hire three archetypes together: a research-oriented engineer who can move models from paper to production, a systems engineer who owns the training and serving stack, and a product-facing engineer who translates capabilities into user-facing features. The first two map to Anthropic's "Research Engineer" and "Distributed Systems" clusters; the third resembles the application-layer hiring that Databricks' vertical directors ultimately serve.

The founding-title prefix (founding ML engineer, founding backend engineer, applied research scientist) signals the company expects candidates to own ambiguous problems end-to-end. Applicants who have shipped a model from data curation through serving infrastructure, and can show the commit history, will match that signal. Those who have only fine-tuned on clean datasets will not.

What the Parsers Actually Read

Parsers from the three dominant providers (Affinda, Sovereign, and Daxtra) process over 2 billion resumes per year. Their models extract and rank entities: skills, tools, frameworks, project scope, measurable outcomes. A 2023 scan of 10,000 resumes found that single-column, ATS-friendly formats received 25 percent more interview callbacks than visually designed resumes with identical content. The layout itself is a signal; multi-column designs, graphics, and non-standard section headers cause parsing errors that drop candidates before a human ever sees the file.

Skills sections carry disproportionate weight. A 2024 LinkedIn internal study showed that profiles with complete skills sections receive 17 times more recruiter views than those without. The parsers map those skills to taxonomies used by downstream ranking models (Eightfold AI's system analyzes over 1 million data points per candidate), so omission is effectively invisibility. Candidates who list "Python" but not "PyTorch," "distributed training," or "model serving" lose matches against roles requiring those specific tokens.

Project descriptions following the STAR format (Situation, Task, Action, Result) score 34 percent higher on average than unstructured answers, even when content quality is similar. The reason is mechanical: language models scoring video or written responses are trained to detect that pattern. They reward explicit problem framing, individual contribution clarity, and quantified impact ("reduced inference latency 40 percent by implementing KV-cache optimization across 8 GPU nodes" beats "optimized model serving").

Public presence functions as a second resume. A 2024 ResumeBuilder study found that candidates with an optimized LinkedIn profile (professional photo, detailed work descriptions, at least 500 connections) were 71 percent more likely to be contacted by recruiters than those with minimal profiles. A 2024 CareerBuilder survey reported that 70 percent of employers screen candidates on social media and 57 percent have eliminated a candidate based on what they found. GitHub repositories with stars, forks, and recent commits; technical blog posts that rank for relevant queries; conference talks or open-source contributions: these are parsed, scored, and fed into the same ranking pipeline.

Cover letters are rarely ingested by parsing layers. Pedigree signals (university brand, previous employer prestige) are weighted but often overridden by skill-match density and project evidence. The system optimizes for recruiter efficiency, not candidate experience. It filters for signal density: how many verified, relevant tokens appear in parseable locations.

How Candidates Clear the Bar

Databricks and Anthropic run multi-stage screens prioritizing demonstrable systems work over credentials alone. Candidates who clear those bars share three tactics.

First, they treat the initial screen as a technical audit, not a conversation. They submit a single-link portfolio (a GitHub repo with a reproducible training run, a benchmark they reproduced, a kernel they optimized) rather than a resume full of project descriptions.

Second, they map their experience to the exact stack in the job posting. Applicants who lead with specific technologies named in the requisition advance; those who use generic descriptors do not.

Third, they prepare for the follow-up coding session by rehearsing the kind of problem the team actually solves: not LeetCode mediums, but systems-level challenges involving concurrency, memory management, or distributed coordination.

The broader shift is toward evidence-based screening. Both companies now require a code review or take-home mirroring production constraints before any human interview. Candidates who anticipate that structure and prepare artifacts in advance (a profiled kernel, a reproducible training script, a failure postmortem they wrote) convert at higher rates. The playbook isn't secret. It's just labor-intensive.

The Funnel Has Already Broken

The multi-stage screen is not an outlier. It is the visible edge of a system-wide recalibration in how AI-native companies separate signal from noise — a recalibration forced by an application flood that has rendered traditional credentials nearly useless as filters.

"It has never been easier to apply for jobs and it has never been harder to find the right candidate," the University of Chicago's Polsky Center concluded in its January 2026 report. Employers see more volume, less differentiation, and skyrocketing fraud and misrepresented experience.

The breakdown of legacy signals is measurable. Stanford's Digital Economy Lab noted that AI usage improves cover letters but makes them less informative signals of worker ability; employers correspondingly shift toward alternative signals such as past reviews. Cui et al. (2025) confirmed the dynamic: cover letters degraded as predictors once AI could write them well. The bachelor's degree, long the default proxy for baseline competence, is losing its signaling power. For the first time in modern history, a bachelor's degree is no longer a reliable path to professional employment, citing a glut of degree holders and declining demand for the credential itself.

Meanwhile, the composition of demand is shifting beneath the surface. Handshake reported a fivefold increase since 2023 in job postings listing AI skills as requirements or qualifications; even internships showed a fourfold increase. The Stanford lab found entry-level hiring in AI-exposed jobs declined 13 percent relative to less-exposed jobs, concentrated among workers aged 22 to 25 in software development, customer service, and clerical work. Older workers saw statistically insignificant impacts. Youth unemployment (ages 16–24) stood at 10.4 percent as of September 2025. Goldman Sachs projects 6 to 7 percent of the U.S. workforce displaced over the next decade.

The occupational mix tells a more nuanced story. Yale's Budget Lab found shifts in occupational composition were well underway before generative AI's release; more recent changes do not appear more pronounced. The share of workers in low, middle, and high AI-exposure groups has stayed stable at roughly 29, 46, and 18 percent. Employment in occupations with high automation-oriented AI usage remains around 70 percent; high augmentation-oriented usage sits near 11 percent. Overall labor market metrics show no discernible disruption since ChatGPT's release 33 months ago, undercutting fears of broad cognitive-labor erosion. But the hiring layer — the gateway — has already transformed.

AI firms are responding by rebuilding the funnel. The emerging recruiting stack, per the Polsky Center's portfolio observations, has three layers: AI for volume and logistics (inbound management, spam filtering, scheduling, candidate communication); video and skills assessments for signal (structured, short-form video responses or practical tests early in the funnel to see the human behind the resume); humans for judgment (managers reviewing top video responses, not keyword scores, making final decisions). A 90-second video answering "Tell us about the most important project you've owned" reveals communication, authenticity, and judgment that AI-generated text cannot.

Deloitte's 2026 Global Human Capital Trends survey found seven in 10 business leaders say their primary competitive strategy over the next three years is speed and nimbleness. Yet organizations taking a tech-focused approach to AI are 1.6 times more likely to miss outsized returns compared to those taking a human-centric approach. Competitive advantage is now less driven by technology differentiation — increasingly ubiquitous — and more by cultivating the human edge. People are not replicable.

The 90-second video is the new filter. A candidate sits down, hits record, and walks through the hardest system they've shipped — the latency budget they met, the deadlock they untangled, the model they took from dirty data to production serving. No prompt engineering survives that frame. The script runs out, and what remains is the work.

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.