180 Applicants, 3% Reach Interview: DeepReach's Robotics Filter

The Screening Gauntlet

DeepReach, a 2026-founded startup capturing first-person stereo video from wearable devices across 475 sites in seven countries, operates in a hiring environment where the average open role attracts 180 applicants and only about 3% reach the interview stage. The U.S. hire rate for technical roles sits at 3.1 percent, the lowest since April 2020, and the broader funnel reflects that pressure: 60 percent of applicants abandon the process before finishing, and senior roles take more than 90 days to fill nearly 40% of the time.

The gauntlet has lengthened across the sector. Interviews per hire have risen 42 percent since 2021, from 14 to 20. Average time to fill is 44 days and costs $4,700 for non-executive roles, up from 33 days and $4,425 four years earlier. Recruiters juggle roughly 20 open requisitions simultaneously, compressing the time they can spend on any single candidate.

AI now powers the first filter. Forty-three percent of organizations used AI for HR tasks in 2025, up from 26 percent a year earlier. LinkedIn found 74 percent of recruiters say AI makes hiring more efficient. Tools like Pin claim to cut resume screening time by 75 percent, interview scheduling by 60 percent, and overall time-to-hire by 31 percent. An Insight Global survey reported 99 percent of hiring managers now use AI somewhere in the process.

The first gate is binary: does the candidate meet the minimum qualifications to be successful? The answer is yes or no. If no, that candidate is no longer considered. Candidates who hold preferred qualifications land on the interview rank list: the shortlist the team actually wants to talk to.

Structured interviews have become the standard. Research shows they predict job performance at roughly twice the rate of unstructured conversations. Companies using them are twice as likely to make successful hires. Two interviews gives you a good idea of who these people are.

Scheduling is a silent killer. A Cronofy survey of 12,000 candidates across seven countries found 42 percent withdrew because scheduling took too long. High-performing employers disposition candidates within three to five days and have 52 percent fewer candidates stuck in extended waiting periods. Sixty-five percent of candidates lose interest after a poor interview experience.

These figures describe the frontier-tech hiring environment DeepReach operates in, a landscape where AI-driven parsing, structured evaluation, and speed-to-disposition separate teams that close talent from those that watch candidates walk.

What DeepReach Is Actually Looking For

DeepReach is hiring across five roles, all anchored to its San Jose headquarters at 2540 N First St #102, with one remote exception. The company organizes technical work around three tracks: Computer Vision, VLA & Robot Learning, and ML Infrastructure. Each position maps to one of them. Roles fall into two tiers: full-time Member of Technical Staff positions that own production systems, and a project-based operator role that feeds the data engine directly.

The most detailed public spec belongs to the Member of Technical Staff, ML Infrastructure role. The baseline: a bachelor's degree or equivalent practical experience in computer science, machine learning, or systems. The real filter is the experience stack: strong engineering background in ML infrastructure, data platforms, training systems, or large-scale experimentation frameworks; fluency in Python and modern ML tooling; hands-on work with distributed training, data pipelines, orchestration, experiment tracking, model evaluation systems, dataset versioning, or GPU job infrastructure. The role owns the infrastructure layer behind data ingestion, curation, training, evaluation, experiment tracking, and deployment workflows: the connective tissue between deployment data, teleoperation data, and model evaluation. The posting emphasizes "fast, pragmatic systems in a startup environment," "strong ownership and debugging skills across infrastructure and ML workflows," and the ability to "go from ambiguous requirements to reliable internal platforms with minimal supervision."

The Member of Technical Staff, Computer Vision and Member of Technical Staff, VLA roles share the onsite requirement and full-time status. VLA — vision-language-action — focuses on robot learning models that ingest multimodal inputs and output control policies. Computer Vision feeds perception pipelines supporting both teleoperation and autonomous deployment. Neither has a public spec as granular as the ML Infrastructure posting, but the company's description of its work, spanning robot deployment, teleoperation, data generation, model training, and evaluation, implies these positions sit at the intersection of research reproduction and production hardening. The founding researcher role advertised on LinkedIn, described as "thinks like a scientist but builds like a startup operator," suggests the VLA track expects candidates who can read papers, reproduce key ideas, and ship internal platforms that accelerate iteration.

The Robotics Research Intern is the only remote role, listed as full-time. The Teleoperation Operator role is project-based, onsite in the Bay Area, and distinct from the engineering tracks. Operators generate the demonstration data that trains and evaluates the models the other roles build. This is not a labeling gig; teleoperation at DeepReach means piloting physical robots through tasks that produce the high-fidelity data the VLA and Computer Vision models learn from. The role sits at the front of the data engine: "how data is collected, filtered, evaluated, and turned into measurable gains on real tasks."

Across all roles, the team structure is onsite by default. Benefits, including health insurance, free food, 401(k), and generous PTO, apply company-wide. The interview process is described as fast, with the company aiming to "move quickly for strong candidates." The common thread: candidates must have already built something that touches the full stack, including data pipelines that serve real models, training runs that complete on physical hardware, and evaluation loops that close the gap between simulation and deployment. DeepReach is not hiring for potential in the abstract; it is hiring for evidence you have already operated in the loop it runs.

The Portfolio Criterion

DeepReach's core product is a hardware-software platform that captures that footage from wearable devices deployed across thousands of real-world work environments, such as warehouses, kitchens, workshops, farms, and clinics, then pipelines that data to frontier robotics companies training world models and manipulation policies. The company frames the bottleneck bluntly: "Physical AI — robots and world models — is bottlenecked on one thing: diverse, real-world data at massive scale showing how humans actually do work in the physical world." That mission shapes what the hiring screen rewards. A resume listing "PyTorch" and "ROS" reads as table stakes; what moves a candidate forward is evidence they have built, debugged, or scaled systems that touch physical hardware in messy, uncontrolled settings.

The research footprint shows DeepReach operating at a scale that demands production-grade infrastructure: 475 devices in the field across seven countries, 100-plus data partners, nearly 150,000 clips collected in three months with volume doubling month over month, and data already in production with multiple frontier model and robotics companies. Candidates who can point to projects involving fleet management of edge devices, real-time video ingestion pipelines, data quality assurance at scale, or simulation-to-real transfer for manipulation policies are speaking the company's language. The ACM-published work on "DeepReach" — a sinusoidal-network-based neural PDE solver for high-dimensional reachability problems — signals that the technical lineage includes formal verification and optimal control, not just deep learning. A portfolio that includes trajectory optimization, reachability analysis, or simulation certifications (NVIDIA Isaac Sim, MuJoCo, or custom physics engines validated against hardware) carries weight because it maps directly to the problem of making robot policies generalize across the "thousands of environments" DeepReach targets.

Data-partner testimonials on the company site, from Quito, Utah, Kanpur, Bangkok, Toronto, and Ha Long, illustrate the operational diversity the platform must support. Each partner runs a distinct business environment with its own lighting, occlusions, workflow variations, and safety constraints. A candidate who has deployed perception or control stacks across multiple sites, handled domain shift without retraining from scratch, or built annotation tooling that non-technical operators can use demonstrates the systems thinking the role requires. The platform's "quality pipeline" and "payments" infrastructure also implies backend engineering challenges: consensus labeling, anomaly detection in streaming video, incentive-compatible reward mechanisms, and audit trails for data provenance. Projects showing end-to-end ownership of such pipelines, especially with human-in-the-loop workflows, stand out more than model-centric notebooks.

Comparable hiring signals at peer companies reinforce the pattern. Zero G Talent's board data for Figure AI lists roles titled "Helix AI Engineer, Training Performance," "Helix AI Engineer, Localization and Mapping," and "Helix AI Engineer, Perception"; these titles emphasize subsystem ownership over generic "ML engineer" labels. Boston Dynamics seeks a "Teleoperations Research Engineer, Atlas" and a "Staff Reinforcement Learning Research Engineer," both roles where a portfolio of hardware experiments, sim-to-real transfer reports, or open-source contributions to manipulation benchmarks (ManiSkill, RoboCasa) would be expected. Zipline's "Operations Ghostbuster" role — a title signaling cross-functional debugging in the field — underscores that frontier robotics hiring filters for people who have shipped in the wild.

No published DeepReach rubric lists specific simulation certifications, GitHub star thresholds, or project templates. The company is young and its public hiring footprint is still forming. But the technical architecture, including wearable capture hardware, distributed data-partner network, quality pipeline, and frontier-model customers, creates an implicit specification: show us a system you built that has withstood real-world deployment, produced data that improved a downstream policy, or solved an infrastructure problem at the edge. That is the portfolio criterion.

How Candidates Clear the Bar

Research on comparable frontier-robotics hiring loops, including DeepMind, Google DeepMind, and the broader cohort building embodied AI, converges on a single number: most candidates who clear the bar spend four to six weeks in focused preparation. That window isn't arbitrary. The interview demands algorithm fluency on par with a Google software engineer, embedded-systems knowledge of a firmware developer, and the mathematical intuition of a physicist. One guide calls it "the most difficult technical interview in existence." Candidates who treat it like a standard coding interview get filtered out in the first technical deep dive.

Technical preparation starts with the core syllabus. Reinforcement learning, robot kinematics, filtering and state estimation, and control theory appear repeatedly as the top tested themes. ROS2 is the industry standard as of 2026; candidates who cannot discuss node lifecycle, DDS tuning, or real-time executor configuration in concrete terms stall early. Simulation tooling matters too: NVIDIA Isaac Sim and Isaac Lab are the environments where policies are trained before they touch hardware. Domain randomization, system identification, and the sim-to-real gap are fair game in system design rounds. RTOS versus Linux, hard versus soft real-time deadlines, and how to run a vision transformer on an Orin while reserving cycles for safety loops are not trivia. They're the daily constraints of the role.

System design preparation separates senior engineers from the rest. The "what" is easy; the "why" is what interviewers score. Guides from both DeepMind and Google DeepMind stress stating assumptions upfront: define constraints, identify the primary goal, then articulate trade-offs between latency, accuracy, and safety before committing to an architecture. Candidates who jump straight to a component diagram without clarifying failure modes — sensor dropout, actuator saturation, compute budget overrun — signal they've never shipped hardware that can hurt people. A recurring pro tip: always ask the interviewer, "What is your fallback strategy when the primary sensor fails?" It shows you understand the first rule of robotics: hardware fails, code must survive.

Project portfolio readiness is non-negotiable. Interviewers expect candidates to discuss past work with extreme precision, not just the "how" but the "why" behind every design choice. If you cannot explain why a specific controller gain was chosen, or why a particular sensor suite was selected over alternatives, you're flagged as lacking depth. Successful candidates prepare two to three quantified examples: "reduced inference latency 30 percent by fusing IMU and visual odometry on the edge," not "worked on sensor fusion." They own failures too; describing a system that crashed, the root cause, and the architectural change that prevented recurrence carries more weight than a flawless success story.

Behavioral rounds use the STAR method, but the robotics twist is technical STAR: Situation and Task frame the engineering constraint, Action details the specific technical decision, Result quantifies the outcome in metrics that matter, such as cycle time, success rate, and safety margin. Practicing answers out loud, recording them, and refining to a two-to-three-minute cadence prevents the nervous speed-talk that buries signal. Mock interviews with a domain peer or coach surface gaps no solo study catches: unclear whiteboarding, missing trade-off articulation, or hand-waving on safety architecture.

Networking and referral leverage change the funnel math. Researching the company for two-plus hours, including products, recent blog posts, leadership interviews, and Glassdoor culture signals, lets a candidate reference specifics in conversation rather than generic enthusiasm. Referrals from current engineers don't bypass the technical bar, but they do get a resume past the initial parsing layer and into a human's hands. In a field where cultural misalignment drives more rejections than skill gaps, a warm introduction that validates collaboration style matters. Preparing five to seven questions that couldn't be answered by thirty seconds of search, such as team roadmap, simulation-to-real pipeline maturity, and how the org handles technical debt in safety-critical paths, signals the same rigor the interview tests.

The night before the onsite, the highest-leverage activity is sleep, not cramming. Morning movement, controlled breathing, and the reminder that interviewers want you to succeed keep cognitive bandwidth available for the ambiguous, open-ended problems that define the day. Candidates who structure their thinking, communicate trade-offs clearly, and demonstrate they've operated where code meets physics are the ones who get the offer.

Why This Approach Matters Now

The physical AI hiring surge isn't theoretical. Deloitte found 58 percent of companies already report at least limited use of physical AI, projected to hit 80 percent within two years, with Asia Pacific leading implementation. Morgan Stanley projects the humanoid robot market growing from tens of millions today. UBS estimates 2 million humanoids in workplaces by 2035, scaling to 300 million by 2050. Amazon has deployed its millionth robot. Waymo has completed over 10 million paid robotaxi rides. Aurora Innovation runs commercial self-driving trucks between Dallas and Houston. The deployment curve is real, and the talent bottleneck is the constraint.

Motion Recruitment shows a 184 percent year-over-year increase in AI job postings, the largest growth segment in tech. Yet 48 percent of organizations planning to expand AI-related roles face a fundamental mismatch: Deloitte identifies insufficient worker skills as the biggest barrier to integrating AI into existing workflows. The Stanford AI Index reports the number of AI scholars moving to the United States has dropped 89 percent since 2017, with the decline accelerating 80 percent in the last year alone. Meanwhile, employment among software developers aged 22–25 has plummeted nearly 20 percent since 2024, and software development job postings have fallen 53 percent from the same baseline. The market is bifurcating: pure software roles contract while physical AI roles expand, but the talent pipeline hasn't reoriented.

Hiring mechanics have shifted in parallel. Eighty-seven percent of companies now use AI-powered platforms to screen resumes, rank candidates, or schedule interviews. Recruiters report 30 percent faster hiring cycles with AI-enabled workflows. But screening criteria have hardened: "Skills Over Degrees" is the stated mantra, with certifications, portfolios, and AI project experience outweighing formal credentials. Yale SOM research finds critical thinking and complex problem-solving rank as the most sought capabilities by a wide margin, followed by adaptability, creativity, and technical and data analysis. Employers "are no longer just looking for workers who can execute tasks. They are looking for those who can exercise reasoning in AI-enabled environments."

The sim-to-real gap drives this. CSET notes the gap between impressive demonstrations in controlled environments and millions of affordable robots acting independently is enormous. Robots succeed in only 12 percent of real household tasks like folding clothing or washing dishes. Deloitte highlights that advances in physics engines, synthetic data generation, and blended virtual-real training should help achieve physical training quality at simulation scale and safety, but only if engineers have actually bridged that gap. Component commoditization and open-source development are reducing entry costs, yet each robotics company pursues a unique approach, leaving the supply chain largely non-standardized. Batteries, motors, sensors, and actuators evolve far more slowly than algorithms. Scalable manufacturing requires patient capital. The engineers who understand these constraints, who've debugged perception stacks on physical robots, tuned actuator controllers across temperature cycles, and orchestrated fleets where comms drop and latency spikes, are the ones who get through the screen.

DeepReach's filter — demanding demonstrable robotics and ML infrastructure experience over generic credentials — mirrors what the entire sector now requires. The companies deploying at scale all screen for the same thing: evidence you've closed the loop between model and machine. The portfolio isn't a nice-to-have; it's the proxy that survives the automated screen and the technical panel. In a market where 63 percent of CFOs plan to increase IT and digital transformation spending, but hiring has slowed to 2010 levels, the candidates who pass are the ones who can show, not tell.

The resume hits the inbox. The clock starts. The screen asks one question: what have you built that survived contact with the physical world?

Working in robotics? Zero G Talent tracks the openings: see every open Zipline role, browse robotics jobs, openings at Boston Dynamics and Figure AI, and the people building the field.