99% of Fortune 500 Firms Use ATS. Roame’s Two-Role Screen Shows What It Means

The Screen Nobody Explains

Roame, a Y Combinator S23 startup that searches over 200 airlines for credit-card-point redemptions, is hiring for two roles, but its screening gate reflects an industry where 99% of Fortune 500 companies use applicant-tracking systems and three in four resumes are rejected before a human ever reads them.

The mismatch is systemic: candidates prepare for culture-fit conversations while screens, increasingly AI-driven, score for quantified problem ownership. Roame's process, opaque as it is, illustrates how frontier companies now filter for engineers who can signal both technical fluency and the judgment to deploy it, a combination that resists keyword optimization.

Eightfold.ai says its AI-enabled screening "adds a structured evaluation layer that helps recruiters gather richer candidate insights earlier in the hiring process," while the ATS remains the system of record. A model may score your resume against a rubric no human has explained to you before a person ever reads it. The criteria, "richer candidate insights," are defined by the company, not published for applicants.

Candidates adapt by reverse-engineering. They read earnings calls, map past work to the company's stated financial goals, and practice structured answer frameworks (STAR, CAR, HERO) until the pattern becomes automatic. They treat the screen as a sales call for a product they've already built: their future output.

The research doesn't show Roame's rubric. It shows an industry where the screen rewards candidates who stop performing "candidate" and start performing "owner of the problem." Whether Roame's two open roles follow this pattern or deviate from it isn't in the public record. What's clear: the gap between what candidates prepare and what screens measure is where most applications die.

Roame's Open Roles

Roame, which bills itself as "Google Flights for your credit card points and miles," sits at an unusual intersection: a consumer travel product built on airline inventory systems never designed to be queried this way. The company searches over 200 airlines to surface award availability for more than 1 million travelers, per its LinkedIn page. That scale — millions of queries against fragmented, rate-limited carrier APIs — is why the engineering roles read more like infrastructure plays than typical application-layer hiring.

Two roles appear consistently across Roame's public postings. The first, on Y Combinator's Work at a Startup board, is a Founding AI Full Stack Engineer. The posting emphasizes an "owner mentality" and the ability to "wear many hats," explicitly calling for familiarity with or interest in "points or aviation/travel." The second, on Vex Dynamics, frames the need as a Lead Full-Stack Engineer tasked with modernizing Roame's Go architecture and overseeing "AI-assisted workflows." That description adds a harder technical floor: expertise in "handling financial systems, distributed engineering, and maintaining high technical standards in a fast-paced environment."

The overlap is intentional. Roame's core product — searching award space across hundreds of programs in near-real time — is a distributed systems problem masquerading as a travel app. The "AI" in the title isn't a chatbot bolt-on; it's the layer that normalizes messy airline data, predicts award availability before it disappears, and ranks redemptions by subjective value. The Lead role's mandate to "oversee AI-assisted workflows" signals the company is moving from prototype models into production pipelines that must stay accurate when a carrier changes its award chart without notice.

Both postings signal the same constraint: Roame needs engineers who can operate across the stack without handoffs. "With Roame, there is no limit to how much responsibility you can take from growth to operations to even product," the Work at a Startup listing states. That breadth is the hiring filter. A specialist who only knows model training hits a wall when the Go service feeding it spikes latency under load. A backend engineer who treats the ML model as a black box won't catch the drift that turns a 90% accurate redemption ranker into a 60% one after a quarter of airline schedule changes.

Only these two named roles appear in the research. Wellfound's board lists "2 jobs in August 2026", a date that may reflect a posting artifact rather than a current count. If additional roles exist, they haven't surfaced in the same public channels with the same specificity. What is documented: a hiring bar blending systems engineering, ML productionization, and domain fluency in loyalty programs, a combination that doesn't exist in standard talent pools.

That specificity is the point. Roame's technical moat isn't a model; it's the pipeline that turns 200 airlines' inconsistent, throttled, poorly documented APIs into a search experience that feels instant. The engineers who keep that pipeline running while improving the intelligence layer are the product. The roles aren't "AI roles" in the industry's current shorthand. They're infrastructure roles that happen to require AI competence, and that distinction is where most candidates misread the screen.

How to Write a Resume the Parser Reads

The screening gauntlet at companies like Roame runs on three mechanical stages: parsing, scoring, ranking. Most resumes die at parsing. The system breaks your document into structured fields — name, contact info, job titles, dates, skills, education — and if your formatting is too complex, this step fails silently. Your data ends up garbled. You never know it happened.

Single-column layout. Standard fonts: Arial, Calibri, Georgia, Helvetica. No tables, columns, text boxes, or decorative graphics. Contact information sits in the body, never in headers or footers; many ATS skip those sections entirely. Save as .docx or a clean, text-based PDF unless the posting specifies otherwise. Before you submit, open the file, select all, copy, and paste into Notepad. If the result looks garbled, out of order, or missing sections, the ATS sees the same mess. Fix it first.

Keyword overlap drives scoring. Jobscan found that resumes hitting 75 percent match against the job description get significantly more callbacks. Push past 100 percent and you trigger spam filters; the system infers you copied the description. Mirror the employer's exact terminology. "Cross-functional collaboration" and "teamwork" are not the same to a parser. Include both the acronym and the spelled-out version: PMP and Project Management Professional. Check spelling conventions across borders; "organisational behaviour" won't match "organizational behavior."

Every bullet point needs three components: Challenge, Action, Result. The CAR framework is what both recruiters and AI models reward. Challenge: the situation in one phrase. Action: a specific verb the parser can extract cleanly. Result: the measurable outcome. This is the part almost every candidate skips. Compare: "Managed customer onboarding process and improved efficiency" versus "Inherited 40-day onboarding backlog. Rebuilt the intake flow and trained the CS team on the new process. Cut time-to-activation from 40 days to 12 across 300-plus accounts." The second one scores.

A dedicated Skills section gives parsers a clean extraction point. List technical tools, platforms, and methodologies relevant to your target role, even if they already appear in your bullet points. Soft skills carry weight now; companies adopting skills-based hiring train screening tools to detect interpersonal signals alongside technical ones. "Stakeholder communication" and "cross-functional leadership" in context are not fluff; they are searchable tokens.

The resume must speak two languages: one for the AI that scans it, another for the human who reads it. Master both and you're not just applying; you're standing out.

Tailoring is non-negotiable. Spray-and-pray fails. Create two to three versions of your resume for similar roles, track results over two to three weeks, then double down on the best performer. Reorder experience so the most relevant work appears first, even if it's older. Use the target job title from the posting in your professional summary near the top; a study found candidates whose resume title matches the target role are 10.6 times more likely to get an interview.

Networking remains the only reliable bypass. After submitting, search the company on LinkedIn. Filter by first- and second-degree connections. Message first-degree contacts directly. Ask second-degree connections for a warm intro via their shared contact. Find the hiring manager by searching "[Department] Manager" or "[Job Title] + Manager", and email them directly using the company's contact structure ([email protected]). Subject line: "Application for [Role] – Following Up Directly." Even if they don't reply, it nudges them to pull your application from the pile.

Some candidates still try the white-text keyword hack, invisible to humans, parsed by bots. Modern ATS detect it and flag the application permanently. Don't. Opting out of AI screening usually means a slow manual review queue that gets touched after the role fills. Optimizing the resume is almost always more effective than avoiding the system.

The parsing and scoring pipeline runs in seconds to minutes for a full batch. That's why automated rejections can arrive the same day you apply. The candidates who advance aren't gaming the system; they've learned its grammar.

Why Every Frontier Screen Now Feels the Same

Roame's opaque screen — where candidates struggle to reverse-engineer what moves an application past the first gate — isn't a company quirk. It's the visible edge of a systemic shift that has turned hiring into an AI-versus-AI standoff.

The Polsky Center documents the result: a flood of lower-signal resumes and employers grading AI against AI rather than evaluating talent. Screens across the industry behave like keyword-matching filters that reward prompt engineering over substance, leaving candidates to guess which phrases unlock the next round.

Geographic concentration reinforces the pattern. Brookings data shows generative AI postings cluster in just ten metro areas, with one in four landing in the Bay Area alone. More recent figures show the U.S. holds three in ten global AI postings in technology, up sharply year over year, while India's share fell. Magnit reports both countries doubled their AI/Automation workforce, yet the most-filled locations among their clients were Los Angeles, Dublin, and Rochester, signaling that even as hiring expands, it clusters around existing talent density and infrastructure. Roame's open roles sit inside this logic: the company is fishing in the same overfished ponds where every other frontier AI firm casts its net.

Demand composition is shifting beneath the surface. AI/Automation fills doubled year over year, from 3% to 6% of total fills, even as overall IT/Tech fills contracted 2%. Within that category, Data Engineering dropped from 46% to 32% of fills while Automation surged from 32% to 44%. Employers are moving beyond pure technical credentials: rising demand for hybrid profiles blends data manipulation and model training with strategic problem-solving and decision-making. Deloitte underscores the pivot: seven in ten leaders say their primary competitive strategy is speed and nimbleness, and organizations taking a human-centric approach to AI are 1.6 times more likely to exceed investment-return expectations than those chasing technology differentiation alone. Screens across the frontier select for candidates who meet the earlier-described criteria, the aforementioned combination easy keyword optimization.

The entry-level pipeline is tightening at the same time. Stanford's Digital Economy Lab found that within firms, entry-level hiring in AI-exposed jobs declined 13% relative to less-exposed roles, with employment drops concentrated among 22-to-25-year-olds in software development, customer service, and clerical work. Yale's Budget Lab confirms the broader labor market has shown "no discernible disruption" 33 months after ChatGPT's release, but the occupational mix is changing faster than in prior tech waves, just not dramatically enough to validate either utopian or apocalyptic forecasts. For applicants, this means the bar for junior roles now often includes production-grade model experience that didn't exist as a hiring criterion three years ago.

Companies are responding with countermeasures that follow a common architecture. The emerging recruiting stack layers AI for volume and logistics, structured video or skills assessments for signal, and human judgment only at the top of the funnel. Deloitte describes a parallel evolution: AI-assisted tools automate defined tasks, AI-augmented models help prioritize candidate assessments, and AI-powered agents begin managing end-to-end processes with minimal human intervention. Interview intelligence platforms now offer real-time feedback to interviewers, attempting to restore signal lost to scripted, AI-generated responses. A multi-stage funnel, resume screen, technical assessment, video response, panel interview, maps cleanly onto this architecture.

Roame's engineers don't just build models; they keep 200 airlines' throttled, inconsistent APIs searchable in near-real time. The screen that filters for that skillset, part systems engineering, part ML productionization, part loyalty-program fluency, is the product. Candidates who crack it aren't gaming a parser; they're proving they can own the pipeline.

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.