Who gets hired and onto which teams
Brett Adcock raised $700 million at a $6 billion valuation in May 2026 to build a universal AI interface that operates the internet through a browser, and eventually through purpose-built devices, then staffed the effort with veterans who have shipped at that intersection before. The Series A, led by Parkway Venture Capital with Nvidia, AMD Ventures, and Qualcomm Ventures participating, answered the funding question. The remaining question is who shows up to build it.
As of May 2026 the company employs roughly 70 people, up from about 45 in March, with a stated target of 100 by mid-year. The hiring plan splits across three pillars Adcock named when the round closed: hardware, product design, and AI research. On the hardware side the roster reads like an Apple alumni directory. David Narajowski and Dave Wilkes, both veterans of product development architecture and audio hardware systems at Apple, lead hardware engineering. Abidur Chowdhury, who spent seven years as an industrial designer on iPhone and Mac products including the recent iPhone Air, runs design as director. Their mandate is to deliver the "bespoke native hardware devices" the company has promised will follow its first multimodal models, expected summer 2026.
The AI research pillar draws from Meta's Superintelligence Lab. Mingbo Ma, Xubo Liu, Xianfeng Rui, Kainan Peng, and Zhihong Lei joined as senior researchers, bringing experience with large-scale multimodal training. They sit alongside modeling and engineering teams that Hark describes as working "hand-in-hand, co-evolving the model alongside the systems that handle the messy reality of the live internet." That phrasing is deliberate: the core product, Handoff, is a computer-use agent that must navigate sites without public APIs (DoorDash, Target, Walmart, OpenTable, Zillow, LinkedIn) by driving a browser like a human. The research organization therefore spans pretraining, post-training, multimodal vision, and multimodal speech, each staffed at the Member of Technical Staff level.
Zero G Talent's board data reflects that structure. Current postings list Mobile iOS Engineer and Mobile Android Engineer roles in San Jose, alongside four distinct Member of Technical Staff tracks: Pretraining, Post-training, Multimodal Vision, and Multimodal Speech, all in San Jose and all priced identically. The board's aggregate salary band for Hark runs across 48 salaried roles, confirming that the company prices every technical track at senior hard-tech levels regardless of discipline.
The research does not mention defense or space contracts; the work described is consumer-facing agentic computing. Candidates should evaluate the role against the actual product roadmap (multimodal models shipping summer 2026, hardware to follow) rather than an assumed mission-critical portfolio.
The team composition signals the execution risk Adcock is underwriting. Hardware leads who have shipped those same products. Researchers who have trained frontier multimodal systems at Meta. A design director who defined the physical language of Apple's latest phone. They are not building a demo; they are building the infrastructure that lets an agent act reliably across the unstructured web, then wrapping that infrastructure in a device that justifies its existence.
What it pays
| Role / Grouping | Salary Range | Median | Source | Roles Count |
|---|---|---|---|---|
| Mobile iOS Engineer, Mobile Android Engineer, MTS Pretraining, MTS Post-training, MTS Multimodal Vision, MTS Multimodal Speech | $180,000–$450,000 | — | Zero G Talent board | 6 |
| All salaried roles (aggregate) | $120,000–$450,000 | $300,000 | Zero G Talent's job board | 48 |
That floor for these positions suggests Hark is not hiring junior contributors into its core AI and mobile teams; the roles map to senior and staff-level scope. The broader board data shows a spread that likely reflects support, operations, or earlier-career engineering roles not appearing in current public listings.
With $800 million in total committed capital (Adcock's $100 million seed plus the $700 million Series A) and a $6 billion post-money valuation, Hark has the capital to compete at the top of the market. The median aligns with what well-funded AI labs pay for engineers who can ship production-grade multimodal systems. The top end is competitive with staff-plus offers at leading AI labs.
Equity is not broken out in the board data, but at a $6 billion valuation the option pool is meaningful. Equity terms are not disclosed in the board data; the high cash ceiling reduces reliance on equity upside to hit competitive total compensation. Geography is a single variable today: every listed role is San Jose. Hark runs its own data center with Nvidia B200 GPUs on site (TechCrunch, May 2026), and the hardware team works alongside research. Remote or hybrid arrangements are not advertised in the current postings. For a candidate evaluating an offer, the band on these six roles is the concrete anchor.
How the hiring process works and what gets candidates through it
Hark has not published a public breakdown of its interview stages, and the available reporting (TechCrunch, Observer, and the company's own blog) does not detail a step-by-step funnel. What the record does show is the caliber of people who have already cleared whatever process exists: the Apple and Meta veterans named above, plus earlier hires from Google, Amazon, Tesla, Figure.AI, and Archer. The company grew from roughly 45 employees in March 2026 to about 70 by May 2026, with the same mid-year target, so the hiring velocity is real and the bar is visible in the names.
A generic interview-prep video (dated August 2025) surfaces in the research corpus, advising candidates to research organizational structure, culture, and interviewer backgrounds; to prepare stories that demonstrate learning from mistakes; to handle resume gaps with concise, rehearsed responses; and to read the room and adjust tone mid-conversation. That advice is not Hark-specific, it carries no attribution to Hark recruiters or hiring managers, and should be treated as general guidance only. No source in the research describes Hark's own resume screen, phone screen, technical assessment, onsite format, or decision criteria.
What can be inferred from the hiring outcomes: the work itself (building a multimodal agentic interface that runs on a dedicated B200 GPU cluster and targets browser-native task execution) demands deep systems fluency across model training, inference optimization, and hardware/software co-design. Candidates who have shipped production models at scale, owned latency-critical paths, or integrated silicon-to-stack pipelines match the profile of the people already inside.
Common disqualifiers are not documented. The research does note one privacy-related constraint the product team is wrestling with: providing a user's life context to an AI assistant without making bystanders uncomfortable or violating privacy. That challenge sits at the intersection of product, policy, and systems design, suggesting that engineers who can reason across those boundaries, not just optimize a loss function, are the ones who stick.
If you're applying, the strongest signal you can send is a portfolio of shipped, measurable work in the exact domains Hark is staffing: the aforementioned training domains, mobile systems at Apple/Google caliber, or hardware bring-up for AI workloads. The board listings are the only public, dated menu of what the company is buying right now.
Where the work happens
Hark operates out of San Jose. Every role on the company's job board lists San Jose as the location. The concentration is deliberate: the company is building a full stack that spans model training, browser-based agent runtimes, and custom consumer hardware, and it wants the teams that own those layers sitting within walking distance of each other.
The most concrete facility the company has disclosed is its own data center running Nvidia B200 GPUs. TechCrunch reported in May 2026 that the 70-person company runs such a facility, a notable capital commitment for a team that size. A month earlier, the Observer noted Hark had "struck a compute deal with Nvidia that will bring thousands of GPUs online next month for pre-training and post-training its systems." The two statements align: the B200 deployment appears to be the first tranche of that larger allocation. For a lab training multimodal models from scratch and serving a browser agent that spins up dedicated virtual machines per request, on-premise compute isn't optional — it's the only way to control latency, cost, and iteration speed at the scale Hark is targeting.
Hardware development is the other half of the facility equation. The presence of Narajowski, Wilkes, and Chowdhury signals a hardware program that goes beyond reference designs. Adcock has said Hark will release its first multimodal models in summer 2026, "followed shortly by hardware devices designed around those systems." That sequence (models first, then purpose-built devices) implies a lab equipped for electrical validation, thermal testing, RF characterization, and the kind of integration work that only happens when silicon, firmware, and mechanical teams share a bench.
The browser agent itself, Handoff, imposes its own infrastructure demands. Every user request spins up a dedicated virtual computer with its own browser, file system, and terminal. At scale, that means managing a fleet of isolated execution environments with sub-second cold-start latency, a systems problem closer to cloud infrastructure than traditional AI serving. The company's benchmark harness (WebTailBench and an internal evaluation suite) also runs continuously against live sites, which requires a test lab that can reproduce the web's hostility: bot challenges, dynamic layouts, session-specific state, and the thousands of edge cases that only appear in production traffic.
Headcount context matters for the physical footprint. The Observer reported 45 employees in March 2026 per that target; TechCrunch cited 70 in May 2026. The first-party board data shows 48 salaried roles currently posted, all in San Jose. That density (researchers, hardware engineers, mobile developers, and infrastructure teams in one building) is the facility strategy. No distributed campuses, no satellite offices. The entire stack lives in one place because the integration points between model, agent runtime, and hardware are too tight to manage across time zones.
What isn't public: square footage, lease terms, whether the data center is on-site or colo'd, the test lab's RF chamber specs, or the hardware bring-up schedule. The company hasn't disclosed those details. But the hiring pattern, the compute disclosure, and the hardware pedigree of the early team converge on a single conclusion: Hark is building a vertically integrated AI lab in San Jose where the distance between a model checkpoint and a prototype device is measured in footsteps, not shipping weeks.
Who thrives here
The clearest signal about who lasts at Hark isn't in a values doc — it's in what the company has chosen to build and how it builds it. In roughly eight months the team has gone from a founder's personal capital to a $700 million Series A, from post-training experiments to a computer-use agent that tops the Online-Mind2Web leaderboard while running at one-tenth the token cost of frontier models. That trajectory only happens when a critical mass of people share a specific constellation of traits.
Full-stack ownership, not specialization. ** ** Hark's explicit strategy is to "own the whole pipeline—foundation models, software systems, hardware and user interfaces—under one roof." The hiring record bears this out: the 70-person roster includes former Apple product-architecture leads alongside senior researchers from Meta's Superintelligence Lab, and a former Apple product executive running design. People who thrive here don't stop at the model boundary; they trace a latency spike from the GPU kernel through the browser runtime to the user's perception of speed. The benchmarks Hark publishes (pass@1 curves across SFT and RL stages, latency broken down by model and browser overhead) are the artifacts of engineers who instrument the entire stack.
Autonomy with a bias toward measurement. ** ** Per the Observer, the company's strategy is to keep the entire stack (models, software, hardware, and interfaces) in one facility. In practice that means a researcher decides whether the next experiment is more SFT data or an RL reward redesign, then proves the call on the leaderboard. The Handoff blog post shows the discipline: "Both stages contribute, and their gains stack, making the fully post-trained model strongest across all three benchmarks." No committee approved that sequence; the team ran the ablations, saw the stacking effect, and shipped. Candidates who need a spec handed to them will not last.
**Comfort in hostile, non-deterministic environments. ** "Unlike a controlled operating system, the internet is a hostile place. Bot blocking, pop-ups, banners, and ads exist specifically to stop, slow, or limit automated agents." That framing from the Handoff launch post is also a filter. The agent must navigate 300 million distinct web domains, fewer than one in a thousand of which offer a public API. Success requires engineers who treat flakiness as a first-class design constraint — building recovery loops, verification steps, and fallback strategies into the model itself rather than papering over them with heuristics. The RL emphasis ("supervised data alone cannot fully teach agents how to deal with this environment and recover from failures") is a direct consequence of that mindset.
**Consumer empathy backed by technical rigor. ** Chowdhury's observation cuts to the hiring profile: "People are really building things to help people make software... but we haven't really seen that for the normal person yet." The team that ships Handoff (ordering food, booking travel, recruiting on LinkedIn) has to care that a saved address auto-fills correctly on DoorDash and that the vision-language model parses the DOM at 1/10th the token cost. That dual lens is rare. It shows up in the compensation bands: The Member of Technical Staff roles in those four areas all share the same range, signaling that the company values the integration layer as much as the model core.
Speed of learning over prior pedigree. ** ** Adcock's own history (Vettery, Archer, Cover, Figure, now Hark) demonstrates a founder who bets on learning rate. The company moved from "building our understanding of how to train computer use agents" to mid-training to pre-training in six months while standing up a B200 data center and securing thousands more GPUs from Nvidia. The headcount plan — 45 to 100 in H1 2026 — is aggressive for a team that already includes talent from Google, Amazon, Tesla, and Meta. People who thrive here treat the half-life of their knowledge as weeks, not years.
**Cost discipline as a design constraint. ** "Superior performance at an order of magnitude lower cost per token" isn't a marketing line; it's a survival metric for a consumer product that must run continuously on user-owned devices. Engineers who instinctively profile token budgets, quantize aggressively, and distill without collapsing capability are the ones who ship. The latency numbers in the Handoff post — model latency per turn with xhigh reasoning, browser latency adding roughly 10 seconds — are the vocabulary of daily standups.
The through-line: Hark selects for engineers who think in systems, measure relentlessly, embrace the messiness of the open web, and refuse to hand off the last mile to someone else. The process filters for it; the work demands it; the benchmarks reward it. When Adcock's first multimodal models ship this summer, the team that built them will already be turning toward the device that wraps them — the same way Figure walked out of the lab, the same way Archer certified. The pattern holds. The next proof is in hardware.
Working in frontier tech? Zero G Talent tracks the openings: see every open Hark role, browse frontier tech jobs, the companies hiring, and the people building the field.