Four Interviews, One Tuesday

What Mercor actually does

Mercor rents domain expertise to AI labs. Frontier labs need doctors, lawyers, bankers, consultants, and PhDs to write training data, grade model outputs, build scoring rubrics, and construct reinforcement learning environments. Mercor recruits those people, vets them with a 20-minute AI video interview, matches them to lab projects, and handles the contracting and payroll.

Business

Expert marketplace for frontier model post-training: RLHF, supervised fine-tuning, rubric writing, and RL environments. They call it the expert service layer for frontier model development.

Customers

AI labs. Publicly named: OpenAI, Anthropic, Meta, Google. They claim all of the top five labs and six of the Magnificent Seven.

How they earn

Markup on contractor hours. Experts average around $85/hour, specialists up to $200, with a reported markup near 35%. Permanent placements carry roughly a 30% fee. Very low fixed cost, since the experts are contractors.

Scale

30,000+ vetted experts across 45+ countries. Paying out more than $1.5M per day to contractors as of October 2025. ARR went from roughly $75M in February 2025 to $450M by October 2025.

Origin

Founded January 2023 by Brendan Foody, Adarsh Hiremath, and Surya Midha, high school debate teammates and Thiel Fellows. Started as a developer recruiting marketplace and pivoted into AI training data.

Rivals

Surge AI, Scale AI, Turing, Micro1, Handshake. Meta's June 2025 investment in Scale cost Scale its OpenAI and Google business, and that scramble is a large part of why Mercor grew so fast.

The connection to your role: the entire asset is a network of people the labs trust to be who they say they are. Identity is not a support function here, it is the thing the product rests on.

Where Mercor is right now

• Scale and money. AI expert-data marketplace. Raised a $350M Series C at $10B in October 2025, quintupling its valuation, and was reported in July 2026 to be discussing a raise near $20B. Described as profitable, with ARR estimates in the $1B to $2B range.
• The role is infrastructure, not login pages. The posting frames it as identity and access across an expert network of roughly 63,000 users spanning thousands of Slack workspaces, enterprise groups, and HR systems, scaling toward 100,000. It names large-scale distributed systems as "the most identifiable signal."
• Why the team exists. Forbes reported in April 2026 that Mercor was grappling with contributor fraud and North Korean IT worker infiltration. Separately, a supply chain compromise of the open source LiteLLM library led to a breach in which ID documents and face and voice biometrics were exposed, with Lapsus$ claiming several terabytes. Meta reportedly paused work pending the investigation.
• Read the through line. Knowing who a contributor actually is, and what they can reach once verified, is now existential to Mercor's enterprise business. That is the team you are interviewing for.

What this round is testing

Fifteen minutes is a filter, not an evaluation. Todd is checking that you are real, that you can say in two sentences why identity infrastructure at this scale interests you, that in-person in SF, NYC, or London is not a problem, and that your timeline works. Lead with distributed systems and correctness, since that is the signal they said they screen on.

Why Mercor

Why this lands: it proves you understand their business model, not just their headcount, and it explains why you want this role rather than any backend job. Do not reach for the breach or the infiltration reporting to make the point. The business-model version says the same thing without sounding like you are auditing them.

Ask Todd

• Is Identity a newly formed team, and what made now the moment to stand it up? Tells you whether you are building from zero or inheriting a system.
• Where is the boundary between this team and trust and safety? Do you own verifying who a contributor is, or only what they can reach once someone else has verified them? The single most important scoping question for this role.
• Operationally, what hurts more today, provisioning access across all those Slack workspaces, or deprovisioning it when someone leaves a project? Shows you know deprovisioning is where the real risk lives.
• What would make it obvious in ninety days that the right person took this job?
• What is the rest of the loop after today, and what is your timeline for a decision?

What Runpod actually does

Runpod rents GPUs to AI developers, billed by the second. Instead of signing a capacity contract, you pick a GPU, launch a container in under a minute, do your training or inference, and shut it down. It is the developer-friendly end of the GPU cloud market.

Business

An AI developer cloud with three shapes: GPU Pods (raw instances you SSH into), Serverless (autoscaling containers that scale to zero, with FlashBoot targeting sub-two-second cold starts), and Instant Clusters (multi-node Infiniband clusters, up to 64 H100s, in minutes).

Customers

Individual AI developers, startups, and teams running inference and fine-tuning. Over 1 million developers, 20+ billion inference requests.

How they earn

Per-second GPU and CPU usage, bare-metal reservations, and serverless execution, with no minimum commitments. Around $120M annualized revenue as of January 2026.

Supply

Two pools. Secure Cloud from enterprise data centers, and Community Cloud aggregating capacity from vetted third-party hosts. 31 regions.

Rivals

CoreWeave and Lambda on the enterprise contract side, Modal and Together on the serverless side, Vast.ai on the marketplace side, and the hyperscalers underneath everyone. Runpod's wedge is granular billing and time-to-first-workload rather than long-term capacity.

The connection to your role: when GPU hourly rates commoditize, the product becomes platform behavior. Cold start time, scheduling, and graceful degradation are the differentiators, which makes reliability engineering the product rather than a cost center.

Where Runpod is right now

• Just raised, and stayed independent. $100M led by Summit Partners at a $1B valuation, announced June 24, 2026, and they say they turned down buyout offers around $500M. Roughly $120M annualized revenue as of January 2026.
• Scale. Over 1 million developers, more than 20 billion inference requests served, 31 regions. They publish developer-experience metrics as the headline: median under an hour from sign-up to first workload, 90%+ first-try deployment success.
• Three products, three very different reliability problems. GPU Pods are raw instances you SSH into. Serverless wraps a model in an autoscaling container that scales from zero, with FlashBoot targeting sub-two-second cold starts. Instant Clusters provision multi-node GPU clusters over Infiniband, up to 64 H100s, in minutes.
• The structural fact that shapes the SRE job. Supply is dual: Secure Cloud from enterprise data centers, and Community Cloud aggregating capacity from vetted third-party hosts. That means running a production SLA on a fleet you do not entirely own or physically control.
• Market pressure. H100 rates have fallen below roughly $1.70/hour, so the differentiation is no longer access to GPUs, it is platform behavior: cold starts, scheduling, billing granularity, and reliability. That is exactly the surface this role sits on.

What the format tends to look like

• Medium-difficulty live coding, usually data transformation or a small systems-flavored problem, with complexity and tradeoff follow-ups rather than algorithm trivia.
• A troubleshooting scenario, likely drawn from their own world: an endpoint's latency spiked, throughput on a pod collapsed, jobs are failing on some nodes and not others.
• They will be grading your method and your narration at least as much as the answer. Think out loud continuously.

Troubleshooting method, say these steps out loud

• Scope the blast radius before touching anything. One customer or many, one region or global, one GPU model or all of them, when did it start, is it getting worse.
• Answer the attribution question first. In a GPU cloud the hard part is usually "is this our infrastructure or the customer's container." Say that out loud and name the signals that separate them.
• Correlate with change. Deploy, config or flag flip, driver or CUDA version rollout, a new node or host joining the pool, image or model weight change, traffic shape change.
• State a hypothesis and the signal that would kill it. This is the step that separates a senior answer from a junior one.
• Bisect. Narrow by region, by host, by GPU SKU, by image, by tenant, by time window until the surface is small enough to read directly.
• Mitigate, then fix. Drain and quarantine the bad host, fail over the endpoint, roll back the driver, shed load. Say what the mitigation costs the customer, because on per-second billing that cost is literal.
• Communicate on a cadence, then write the follow-up. What health check or alert would have caught this before a customer did.

Vocabulary that shows you know this domain

If a GPU-specific failure comes up, these are the real ones, and using the right names buys you a lot of credibility fast:

The cold-start breakdown is a particularly good thing to reason about unprompted: container start, image pull, model weight load into VRAM, and first-token warmup are four separate costs with four separate fixes, and conflating them is the usual mistake.

Your own material to reach for

The interval-sweep work you did on findDeployableWindows travels well here, both as a live-coding warm start and as a bug story, because it contains two of the most common silent-corruption classes in production:

• Type coercion at the input boundary. The string versus integer parsing bug in the CSV input. Generalize it: untrusted input crossing a boundary with no schema, failing quietly rather than loudly.
• Timezone and boundary arithmetic. The UTC offset conversion with week-boundary wraparound. Off-by-one at a wrap boundary is the bug that passes every test and fails at 11pm Sunday in one region only.
• Separating mechanism from policy. Post-filtering by lead time, minimum window length, and result count kept the core sweep correct and the policy layered on top. That framing maps almost directly onto scheduling and capacity work.

Why Runpod

Why this lands: Chen is an engineer, not a recruiter. Naming a specific technical constraint of their business is worth more than any amount of enthusiasm. It also sets up the questions you plan to ask, so the conversation flows.

Ask Chen

• Where does the reliability boundary sit with Community Cloud? How do you detect and quarantine a bad host on hardware you do not physically control? This is the most interesting structural problem in their business, and almost no candidate will ask it.
• At 3am, how do you tell "our node is broken" from "the customer's container is broken"? What signal settles it?
• What is the cold start budget on Serverless in practice, and what dominates it now, image pull or weight load into VRAM?
• What are the top few failure modes you actually page on, and how much of it is GPU hardware versus orchestration?
• With per-second billing and no minimum commitment, what happens during a capacity crunch? Is there preemption, and who absorbs the failure?
• Thirty-one regions and a small team. What does on-call look like, and is it follow-the-sun yet?
• Is this role closer to platform engineering or to classic on-call SRE, and how much of the week is project work versus operational load?

What Chime actually does

Chime is a neobank: banking through an app, with no branches, no monthly fees, and no overdraft fees. Importantly, Chime is not itself a bank. The Bancorp Bank and Stride Bank hold the charter and the deposits. Chime owns the app, the member relationship, and, since late 2025, the transaction processing underneath.

Business

Fee-free consumer banking aimed at everyday Americans, historically the $35K to $65K income band, now deliberately moving upmarket.

Customers

10.2 million active members. The $75K+ income segment is now the fastest growing, which is what Chime Prime is built for.

How they earn

Roughly 80% from interchange, the slice of the fee a merchant pays whenever a member swipes their Visa debit card. That makes it a volume and engagement business: revenue rises when members use Chime as their primary account, not when they hold a balance.

Products

Checking and savings, SpotMe fee-free overdraft, Credit Builder secured card, Early Pay, MyPay earned wage access (past a $400M run rate), and Chime Prime. Coming: investing accounts, joint accounts, and the Jade AI copilot.

Origin

Founded 2012 by Chris Britt (previously Green Dot and Visa) and Ryan King (previously Plaxo). Now public on NASDAQ as CHYM.

Rivals

Cash App, SoFi, Dave, Varo, Current, and the large banks. Regulatory scrutiny of fintech partner-bank models is a live risk across the category.

Why this shapes the engineering: because revenue is per transaction, every point of processing cost is margin. That is exactly why ChimeCore, bringing processing in house, moved the numbers as much as it did.

Where Chime is right now

• Turned the profitability corner. Q1 2026 was Chime's first quarter of positive GAAP EPS, on 25% revenue growth and 13 points of adjusted EBITDA margin expansion. 10.2 million active members, transaction margins around 76%.
• ChimeCore is the engineering story. They finished migrating off third-party transaction processors onto their own core in Q4 2025. Analysts put the processing cost reduction near 60% and a long-term gross margin target around 90%. This is a vertical integration of the money movement stack, and it is what leadership credits for the margin expansion.
• Product direction. MyPay, their earned wage access product, is past a $400M run rate at 62% margins. Chime Prime targets higher-income members, the fastest growing segment. Coming: investing accounts, joint accounts, and a "Jade" AI copilot.
• Worth knowing before you walk in. On the Q1 call they said AI-assisted development accounted for 84% of code shipped in March. That is an unusual thing for a public fintech to state, and it likely shapes what they screen for.

What this round is testing

A recruiter screen is logistics and motivation. Have three things ready to say without hesitating: why you are looking, a two-minute version of your six years, and a compensation number. Since Chime is public, bands are more rigid than at a startup and equity is RSUs against a real share price, so the comp conversation can be concrete rather than hand-wavy. Ask for the band rather than anchoring first.

Why Chime

Calibrate for the audience: this is a recruiter, so keep it to about thirty seconds. They are screening for whether you have a coherent reason at all, not for depth. Save the ChimeCore detail for the hiring manager round unless they ask a follow-up, then you have somewhere to go.

Ask the recruiter

• Which org does this req sit in, ChimeCore and platform, member-facing product, or risk and fraud? These are very different jobs. Find out before the loop, not during it.
• Is the role mapped to a specific team already, or is this a generalist pipeline with team matching at the end?
• What level is this mapped to, and what is the band? How do they map six years of experience?
• What does the full loop look like, how many rounds, and over what timeline?
• What is the location and in-office expectation for this req?
• Given how much of the codebase is now AI-assisted, has that changed what the interviews screen for? Signals you actually read their earnings call, which almost no candidate does.

What Kong actually does

Kong builds the layer that sits in front of APIs. Every request from a client to a service passes through the gateway, which handles authentication, rate limiting, routing, retries, and observability, so that hundreds of individual services do not each reimplement the same concerns. Kong's bet now is that AI traffic needs the same choke point.

Business

Open core. Kong Gateway is free and open source, built on Lua and OpenResty over nginx. Revenue comes from Kong Enterprise and Konnect, the SaaS control plane that manages fleets of gateways.

Products

Kong Gateway, Konnect, Kuma and Kong Mesh (service mesh), Insomnia (API design and testing, acquired 2019), AI Gateway, and Agent Gateway.

Customers

Enterprises running microservices at scale, where API sprawl and governance have become the problem.

The AI bet

The same choke-point argument applied to LLM, MCP, and agent-to-agent traffic. If every AI call routes through a gateway, that is where you put cost attribution, access control, and audit. Agent Gateway shipped in April 2026 as the concrete version of this.

Origin

Began in Milan around 2009 as Mashape, an API marketplace. Released Kong as open source in 2015, sold the marketplace to RapidAPI, and rebranded as Kong Inc in 2017. Founders Augusto Marietti (CEO), Marco Palladino (CTO), and Michele Zonca.

Rivals

Apigee (Google), MuleSoft (Salesforce), AWS API Gateway, Tyk, Solo.io, and teams rolling their own on Envoy. On the AI side, a fast-filling field of LLM gateway startups.

Why this shapes the engineering: a data plane sitting in the request path for all of a customer's traffic is an unforgiving place to build. Latency budgets are in single-digit milliseconds and a bad release is everyone's outage at once.

Where Kong is right now

• The core business. Kong Gateway, the open source API gateway built on Lua and OpenResty over nginx, plus Konnect as the SaaS control plane, service mesh, and Insomnia. Raised $175M at a $2B valuation in a Series E, and is frequently discussed as an IPO candidate.
• The pivot they are betting on. Kong is repositioning from API connectivity to AI connectivity. In April 2026 they shipped Agent Gateway in AI Gateway 3.14, governing agent-to-agent traffic alongside LLM and MCP traffic, with unified observability, token and cost tracking, access control, and audit logging of A2A conversations. Their claim is that they are the only gateway covering LLM, MCP, and A2A from one control plane.
• Recognition. Won a 2026 API Award for best in API infrastructure.
• Read on Arun. A site leader growing an engineering org, out of a large enterprise SaaS background running Jira Service Management. Expect weight on ownership, operating discipline, incident and quality practice, and working well across time zones, more than on puzzle solving.

What this round is testing

Hiring manager rounds decide two things: can you own a problem end to end without being managed, and will you function in this specific team's operating model. With a site leader, the distributed-team question is real and not a formality. Have one story ready about owning something across a time-zone split, and one about a reliability or quality practice you introduced rather than inherited.

Why Kong

With a hiring manager, answer the second half too. Arun will also want to know why his team. Follow up with a version of "I don't know yet which surface this role covers, and that's actually my first question," then ask the charter question. Turning "why us" into a two-way scoping conversation is exactly what a hiring manager round is for.

Ask Arun

• What is this team's charter, the data plane and core gateway, the Konnect control plane, or the AI and Agent Gateway line?
• How is scope divided between the US and India sites, and where would this role sit in that split? Directly relevant given he leads the India site. Asking it shows you understand what his job is.
• Coming from Jira Service Management, which operating practices did you bring to Kong, and which did you deliberately leave behind? Managers rarely get asked this and it usually gets a real answer.
• Agent Gateway is only a few months old. How much of the roadmap right now is pulled by customers versus a bet on where agentic traffic goes?
• The data plane is Lua and OpenResty. Where is Rust or Wasm actually landing in the stack, and where has it not been worth it?
• What is the hardest constraint on the AI gateway in practice, latency budget, multi-tenancy, or provider heterogeneity?
• How do you measure whether an engineer on your team is doing well?