Who gets hired and onto which teams
The rebrand from Captions to Mirage in September 2025 signaled more than a name change. It marked a shift from a creator-tool company into an AI research lab building multimodal foundational models for short-form video. That pivot reshaped who the company needs: Mirage hires for technical depth and cross-functional impact, offering competitive pay and a structured interview process that screens for both skill and cultural fit.
As of late 2025, hiring centers on three pillars. Research engineers sit at the core: roles like Research Engineer, Agentic Systems and Research Engineer, Generative Video appear on the board with bands of $175k–$300k, reflecting the premium placed on publications-grade ML experience and hands-on model development. Around them, software engineers build the infrastructure that turns research into product: Software Engineer, Agents; Software Engineer, Backend; and Software Engineer, iOS all share the same $175k–$300k band (iOS tops out at $275k). A single Group Product Manager role at $200k–$325k anchors the product side.
Twelve salaried roles are listed, all based in Union Square, New York City, a deliberate colocation choice for a lab that treats research and engineering as a single conversation. Third-party listings on LinkedIn, Indeed, and Prosple also surface internship and new-grad programs, but the board's salary bands and title seniority show the near-term hiring weight sits with experienced ICs who can operate autonomously in a fast-moving research environment. The median band of $275k across roles reinforces that Mirage competes for the same talent pool as frontier AI labs.
What distinguishes the team composition is the explicit split between agentic systems — orchestration, tool use, multi-step reasoning — and generative video proper. That architecture mirrors the company's stated focus on "frontier AI research and models" for TikTok, Reels, and Shorts formats. Candidates who land interviews typically show depth in one of those two lanes plus enough systems fluency to collaborate across the boundary.
The hiring mix also reflects a company still small enough that every hire changes the team's center of gravity. With roughly a dozen open roles, each addition carries outsized influence on culture and technical direction. That dynamic favors candidates who have owned ambiguous problems end-to-end (shipping a model from experiment to production, or designing an evaluation framework the org adopts) over specialists who optimize a single metric in isolation.
The research-engineer roles don't just ask for transformer expertise; they look for experience wrestling with video data, temporal consistency, and the compute constraints of short-form inference. The software-engineer roles similarly weight systems work: serving stacks, agent orchestration frameworks, mobile deployment. Generalist ML engineers who lack either the research publication record or the production systems track record tend to stall in screening.
For candidates, the takeaway is structural: Mirage hires in cohorts that balance research ambition with shipping discipline. The team composition today — heavy on research engineers, light on management layers, anchored by a single senior PM — is the org chart of a lab that intends to publish, open-source, and ship models in tight cycles. The next hiring wave will likely mirror the same ratio unless the product surface area expands dramatically.
Compensation: what the numbers show
The board's live postings show a tight cluster: every engineering and research role currently listed falls between $175,000 and $300,000 base, with the Group Product Manager role stretching to $325,000.
| Role | Base salary range (USD/year) | Location |
|---|---|---|
| Group Product Manager | 200,000 – 325,000 | Union Square, NYC |
| Software Engineer, Agents | 175,000 – 300,000 | Union Square, NYC |
| Software Engineer, Backend | 175,000 – 300,000 | Union Square, NYC |
| Research Engineer, Agentic Systems | 175,000 – 300,000 | Union Square, NYC |
| Research Engineer, Generative Video | 175,000 – 300,000 | Union Square, NYC |
| Software Engineer, iOS | 175,000 – 275,000 | Union Square, NYC |
The spread within each band is wide by design; it leaves room for level, scope, and negotiation without creating separate titles for every increment. A backend engineer joining at the top of the range is effectively being paid at staff-equivalent levels; the same title at the bottom reads as a strong senior hire. That flexibility matters because Mirage's interview process evaluates depth first, then slots candidates into the appropriate slice of the band.
Equity data isn't published on the board listings. Candidates should expect option grants that vest over four years with a one-year cliff (the industry default) and should ask for the current 409A valuation and total outstanding shares to calculate percentage ownership. The company's funding history (over $100 million raised at a $500 million valuation, according to TechCrunch) suggests the strike price will be low, but the float is still small enough that early grants can represent meaningful upside if the agentic-video thesis pays off.
Benefits fill out the package. The Wellfound culture page and company materials enumerate flexible paid time off, paid parental leave, a monthly wellness stipend, laptop of choice plus home-office setup budget, co-working allowance, and life insurance.
One tension worth noting: third-party aggregators show wildly divergent figures. Levels.fyi reports a median of ₹733,090 (roughly $8,800) for a Marketing role, almost certainly a different Mirage entity or a misclassified data point. Glassdoor shows 15 anonymous salaries across 15 jobs but doesn't break them out by function. Those numbers may reflect international hires, contract roles, or the casino-hotel entity that shares the name. The board data (first-party, dated, and tied to specific NYC-based technical roles) is the only internally consistent set.
Negotiation leverage comes from knowing the band cold. The company publishes ranges on every listing, so a candidate who can demonstrate staff-level impact on agentic systems or generative video has a credible case for the top quartile. Bring competing offers, but also bring evidence: open-source contributions, papers, or production systems you've owned that map directly to Mirage's roadmap. The band is real; where you land in it is the conversation.
How the hiring process works
Mirage runs its entire interview loop in person at its Union Square headquarters. There is no hybrid or remote option for the assessment stages: candidates fly in or commute to a single day of high-intensity, face-to-face sessions designed to mirror the collaborative, fast-paced rhythm of its research teams. The process tests two things simultaneously: technical fluency at the frontier of generative models and agentic systems, and the ability to operate inside an R&D environment where requirements shift daily and latency budgets are measured in milliseconds.
The loop opens with a deep-dive technical discussion that moves quickly from fundamental concepts (diffusion model architectures, transformer attention mechanics, CUDA memory hierarchies) into specific, hands-on challenges. Interviewers expect candidates to walk through research papers or projects they have authored, explaining not just the "what" but the "how" of implementation choices: why a particular sampling strategy, where the memory bottleneck lived, how they profiled a kernel to shave 5 ms of latency. The culture emphasizes transparency and direct, evidence-based communication; vague answers or hand-waving are flagged immediately.
Whiteboarding sessions are not algorithmic puzzles divorced from the product. They are practical engineering exercises drawn from the team's current work: optimizing a video generation pipeline, designing an agent orchestration layer, debugging a distributed training run. Candidates who thrive treat the whiteboard as a collaborative design tool, not a performance stage. The interviewers are future peers; they are evaluating whether they want to debug a production incident with this person at 2 a.m.
Screening criteria are explicit about what differentiates a good candidate from a great one. A strong applicant knows the theory. A great candidate has an opinion on what works and what doesn't, grounded in their own experimental history, and can defend those choices with data. The "builder's mindset" the hiring guides describe — comfort writing a research paper and profiling a CUDA kernel is the consistent thread across research engineer and research scientist roles. Candidates who have only published or only shipped tend to struggle in the cross-functional segments where the loop tests translation between research insight and production reality.
Preparation that works: bring a concrete project where you made a non-obvious trade-off, quantify the impact, and be ready to argue the counterfactual. Preparation that fails: memorizing model architectures without having touched the training loop, or treating the interview as a lecture rather than a technical dialogue. The process selects for people who have already operated at the intersection of research and engineering, because that is where Mirage lives every day.
Where the work happens
Mirage's technical roles are anchored in Union Square, New York City, according to every recent posting on the Zero G Talent board. The six open positions (Group Product Manager, Software Engineer (Agents, Backend, iOS), and Research Engineer (Agentic Systems, Generative Video)) all list that location.
The product itself shapes the workspace. Mirage "gives AI agents one virtual filesystem and one virtual terminal for S3, Google Drive, Slack, Gmail, Redis, Postgres and SSH, so the unix tools they already know compose across services," per the project's documentation at mirage-bay.vercel.app. The runtime executes in-process: ws.execute(...) parses and dispatches commands without mounting anything on the host. FUSE is an optional surface for editors, language servers, and ripgrep to see the workspace. That architecture means engineers dogfood the platform daily, the same virtual filesystem they ship runs their local development loop.
No public floor plans, lab photos, or lease details appear in the research. The company's technical docs (docs.mirage.strukto.ai) describe architecture, not square footage. What's verifiable: the team works where the postings say, building a tool that turns their own machines into the product's first test environment.
Who thrives here
The review signal from Mirage's roughly 119 combined Glassdoor and Indeed entries paints a consistent picture: people who stay and advance tend to share a specific cluster of traits. The most cited positive is a friendly, teamwork-oriented atmosphere with managers described as "flexible", a phrase that appears repeatedly across both platforms. In a company building full-stack foundation models for video generation, that flexibility matters. Research engineers and software engineers working on agentic systems or generative video models operate at the bleeding edge of a field where requirements shift weekly.
Leadership gets credit for promoting growth and learning, which aligns with the hiring mix: research-engineer roles spanning agentic systems, generative video, and backend infrastructure, plus a Group Product Manager seat. The compensation bands ($175k–$300k for individual contributors, $200k–$325k for the GPM) signal that Mirage pays for depth, not just headcount.
The friction points in reviews are equally instructive. Micromanagement and favoritism surface as recurring complaints. High performers who thrive here tend to be low-ego collaborators who document decisions aggressively, communicate progress visibly, and build their own feedback loops rather than waiting for a manager to supply them. They also tend to be comfortable in the office five days a week, every current posting lists Union Square, New York City, with no remote option mentioned.
The LinkedIn description ("Go from idea to video with frontier AI") doubles as a filter. People who thrive treat video generation as a systems problem spanning data, model architecture, inference optimization, and product integration. Specialists who only want to tune loss functions or only want to build UIs tend to frustrate the cross-functional rhythm the review data describes.
One tension worth flagging: the review sample is small (19 Glassdoor, 100 Indeed) and undated in the digest, so the culture snapshot could lag the current team by 6–12 months. But the consistency across platforms (flexible managers, growth orientation, in-person collaboration, occasional micromanagement) suggests a stable cultural core. If you need autonomy within a clear technical direction, value visible iteration over polished presentations, and want your commits to appear in a product used by millions (the LinkedIn bio cites "over 20 million"), Mirage's profile matches. If you optimize for process comfort or remote flexibility, the signal says look elsewhere.
The rebrand that opened this story — Captions to Mirage, tool to lab — is still unfolding. The next time the board updates, the roles may shift, the bands may widen, but the filter will hold: they hire people who have already built the thing they're being asked to scale.
Working in AI? Zero G Talent tracks the openings: see every open Mirage role, browse AI jobs, the companies hiring, and the people building the field.