Forward-Deployed Engineer

Synthesized from twenty live postings across two packs, read August 9 to 10, 2026: a customer-facing set (OpenAI, Anthropic, Databricks, Google Cloud, AWS, NVIDIA, Salesforce, Vercel, Cresta, Glean) and a full-stack build set (OpenAI, Palantir, Anthropic, Google Cloud, AWS Professional Services, Scale AI, Databricks, Baseten, Vercel, Cognition). OpenAI, Anthropic, Google, AWS, Scale, Databricks and Vercel run both under different titles, and do not agree which one to pay more.

You are the person a customer talks to when they want an AI system and do not know what they want. On the customer-facing ladder you write code when it is the fastest way to make the point, but you are measured on whether the customer's problem got smaller. On the full-stack ladder you write production code inside the customer's cloud account, against systems you cannot change, on data you did not clean, under a security review you did not schedule.

You work with the customer's executive sponsor (non-technical, budget owner), their IT and security function, their frontline users, and, on the build ladder, their engineers in their repo and their code review. You work alongside your own account team; six of ten customer-facing postings pair you with a seller and Databricks asks for a quota track record outright. Experience floors run from 3+ years (Cresta) to 10 (Google Customer Engineer III). Travel of 25 to 50% is stated in six of ten. Python plus one of TypeScript, Go, Java or SQL is the floor.

The two ladders share a customer and split on where the hardest skill lives. Ten shifts show up across both packs and matter for both jobs.

Every named technology in twenty postings is a substrate: a language, a cloud, a framework the vendor sells, or a system you integrate with. Not one names an agent framework, vector database, or eval platform by brand. Six name AWS, Azure or GCP, but as an environment you will be dropped into. “I'd deploy it in their VPC” earns nothing. “Everything runs in their account, so the model call is the only thing crossing a boundary, and that is the sentence I have to get past their CISO” earns everything.

So the rule for this hub: award nothing for naming a framework or a cloud; require reasoning about the substrate. “I'd cache the contract-rate lookup because the API is limited to 60 a minute and my peak hour needs 400” is signal. “I'd use Redis” is not.

• OpenAI

  $162K to $325K + equity

  Three ladders under one function: FDE (scope work, sequence delivery), FDSWE (design, architect, and develop full-stack solutions), and Solutions Architect. The clearest case of the market hiring two people.

  Both ladders

• Anthropic

  $200K to $320K

  FDE, Applied AI works within customer systems to build production applications. Helping customers develop their own evaluation frameworks is a stated deliverable.

  Both ladders

• Databricks

  $139.7K to $301.3K

  Named organizational split: pre-sales Solutions Architect and post-sale Delivery Solutions Architect, with a $79K gap at the top. Asks for a quota track record.

  Both ladders

• Google Cloud

  $174K to $253K + bonus + equity

  Customer Engineer leads MVP development; FDE, Generative AI is live in 23 US metros at once.

  Both ladders

• AWS

  $131.3K to $239K

  Specialist Solutions Architect (agentic AI) runs hackathons and reference implementations; Professional Services Delivery Consultant ships inside the customer account. The only employer verifiably above 20 open reqs.

  Both ladders

• Scale AI

  $138K to $292.6K

  Public-sector FDSWE with the clearest location-banded range in the sample and TS/SCI clearance tracks.

  Build

• Palantir

  $135K to $200K

  The origin of the title. Forward Deployed Software Engineer, New York; Levels.fyi total-comp headlines run higher but are secondary source.

  Build

• Baseten

  $165K to $330K + equity

  Highest ceiling in the build sample after OpenAI, from a company with nine total open roles.

  Build

• Vercel

  $137K to $284K OTE

  Solutions Architect is paid on OTE, which tells you how far toward pre-sales it sits; FDE owns migrations and builds.

  Both ladders

• Cresta

  $185K to $235K + bonus + equity

  3+ years against a $185K floor: the most candidate-accessible high-paying req in the pack. Demos, POCs, trusted advisor, plus Python and Go.

  Customer-facing

• Glean

  $160K to $270K

  Founding FDE: 0-to-1 build experience and technical partner to the C-suite in the same posting.

  Both, in one req

• NVIDIA

  $152K to $287.5K

  Senior AI Solutions Architect; the one posting in twenty that asks for a graduate degree, which we flag and do not replicate.

  Customer-facing

Provn is not affiliated with any employer listed. Descriptions summarize each company's public job posting as read in August 2026.

Customer-facing ladder

Scoping, expectation-setting, diagnosis, and architecture defense. No executable output is required at this ladder by design; the scarce skill here is judgment, and the video is a work sample, not a walkthrough.

1. Practice20 min total

   The Ask Behind the Ask

   Meridian's COO wants “AI for the quote desk.” Convert the stated want into a bounded first deliverable and name the dependency you do not control.

   Isolates: Can you convert a customer's stated want into a bounded first deliverable, and name the dependency you don't control? 3 to 4 min video.

2. Easy28 min total

   Eighty-Seven Percent

   The extractor is right 87% of the time. Tell a non-technical buyer what that means, in a number, with a consequence, and a written non-goals list.

   Isolates: Can you tell a non-technical buyer what a probabilistic system will and won't do, in a number, with a consequence, and a written non-goals list? 4 to 5 min video.

3. Medium40 min total

   The Pilot That's Stalling

   Six weeks in, usage is falling and the sponsor is nervous. Diagnose the stall, re-scope, and deliver the bad news in a way that keeps the account.

   Isolates: Can you diagnose a degrading deployment, re-scope it, and tell the customer the bad news in a way that keeps the account? 5 to 6 min video.

4. Hard40 min total

   The Room Where They Decide

   Two architectures, one budget, a non-technical buyer, and a hostile IT lead. Choose on cost, time and risk, and defend it.

   Isolates: Can you choose between two architectures on cost, time and risk, and defend it to a non-technical buyer and a hostile IT lead? 6 to 8 min video.

Full-stack ladder

Building inside Meridian's AWS account: the code, the integration, the eval on their data, the guardrail, the runbook. Pack recommendation: Practice and Easy ship in September; Medium and Hard open in October.

1. Practice20 min total

   What Breaks Behind Their Firewall

   Code that runs clean on your laptop. Name what fails inside the customer's environment: the egress path, the rate limit, the stale replica, the credentials.

   Isolates: Can you read code running in a customer's environment and name what will fail there but not on your laptop? 3 to 4 min video.

2. Easy28 min total

   The Integration You Don't Control

   A rate limit of 60 a minute, a four-hour replication lag, and a hand-maintained rate file. Write the fallback path and the pre-delivery check.

   Isolates: Can you write the fallback path against a rate limit, a replication lag, and a hand-maintained file? 4 to 5 min video.

3. Medium40 min total

   Blind in Their VPCopens October

   A production failure, logs and telemetry only, no debugger, no deploy window. Diagnose it and say what you would instrument next time.

   Isolates: Can you diagnose a production failure from logs and telemetry alone, with no debugger and no deploy? 5 to 6 min video.

4. Hard40 min total

   Ship It in Their Cloudopens October

   Stand up a production deployment under someone else's IAM, with the eval, the guardrail, and the runbook for the team that did not write it.

   Isolates: Can you stand up a production deployment under someone else's IAM, with the eval, the guardrail and the runbook? 6 to 8 min video.

Both ladders share the spine. On the customer-facing ladder, video carries 25% at every tier because it is a work sample: eight of ten postings make communicating with an executive or non-technical buyer a named responsibility, so the explanation is the job. Commercial and stakeholder judgment is weighted 20 at Easy, the one tier where it outranks the AI-systems dimension, because Eighty-Seven Percent is a commercial-judgment exercise in a technical costume.

What Transformative looks like

Where two figures appear above, the first is the customer-facing ladder and the second is the full-stack ladder. On Eighty-Seven Percent, Transformative states the accuracy figure to the buyer in their terms, attaches a consequence to the other 13%, names who catches it and how, and writes a non-goals list the customer could hold you to. Adoptive states the number and a consequence and commits to a bounded scope; that is the advance bar. Capable gives the number but hedges the consequence into “we'll monitor it.” Below Capable gives the number but attaches no consequence and no non-goals list to it. Not Yet promises the system will work, or refuses to give a number at all.

On this hub the communication standard is audience fit and sequencing: did the listener get the conclusion before the reasoning, was it pitched at the person it was aimed at, did the technical argument survive translation. Never polish, accent, pace, or filler words.

1. A customer says they want “AI for the quote desk.” What do you do in the first meeting?

   Find the ask behind the ask. Which part of the desk's day costs the most, what data exists, what system owns it, and what the customer would accept as proof it worked. The strong answer leaves the room with a bounded first deliverable and the one dependency you do not control written down.

   Where it comes from: Nine of ten customer-facing postings name discovery and scoping; Glean's own posting warns candidates the decisive round is decomposition.

2. The extractor is 87% accurate. How do you explain that to a COO?

   In their unit, with a consequence. Thirteen of every hundred quotes need a person; here is who that person is, here is what it costs, here is what the 87 buys them, and here is what we are not promising. The strong answer treats the number as the start of the scope, not a disclaimer at the end.

   Where it comes from: Anthropic's excluded SA posting: help customers develop evaluation frameworks to measure performance. Five of ten make trusted advisor a responsibility.

3. Your code works in staging. It fails inside the customer's VPC. Where do you look first?

   The network, not the model. Egress rules, the proxy and its CA bundle, the private endpoint, IAM permission boundaries, and whether you are reading a stale replica. The strong answer lists what it would have instrumented before the first deploy because there is no debugger in their account.

   Where it comes from: Scale, AWS and Vercel require deployment into customer-owned infrastructure; observability is something the engineer creates, not consumes.

4. The customer's IT lead did not choose you and controls whether your container runs. How do you work with them?

   Design for their review instead of around it. Image scanning, approved-vendor lists, permission boundaries and change windows are scheduling dependencies with a probability attached, so surface them in week one and make the IT lead's constraints part of the architecture argument.

   Where it comes from: Eight of ten customer-facing postings name executive or non-technical stakeholder communication; the build pack names security review at three.

5. Two architectures, one budget. How do you decide and how do you defend it?

   Cost, time and risk, in that order the buyer cares about them, with the losing option's best argument stated honestly. The strong answer names what it would do if the chosen architecture's riskiest assumption fails.

   Where it comes from: Hard tier, both ladders; six of ten name codifying one-off work into reusable patterns.

6. Tell me about a time you caught an AI tool being wrong on a customer engagement, and what you did about it.

   Name the specific moment: what the tool produced, why it was wrong for this customer's environment, and what you shipped or said instead. Generic answers about “always verifying” score poorly. This is also the mandatory AI question in every Proving Ground video walkthrough.

   Where it comes from: Every posting in both packs assumes AI-tool fluency, but none of them can screen for judgment about when to override the tool; the video makes that judgment visible.