Forward Deployed Engineers: Palantir, Lyra, the AI $1M hype, and Solana

Every company can buy roughly the same frontier model. Two of them will get opposite results. The difference is almost never the model card. It is whether someone sat inside the messy workflow, decided what should stay deterministic code, what needs judgment, what needs a human, then shipped a system operators will actually use — with evals, permissions, and a rollback path.

That person is a Forward Deployed Engineer (FDE). The title exploded in AI hiring in 2025–2026. The idea is older: Palantir built its delivery model around engineers who embed with customers and reshape software around operational reality. Studios like Lyra Technologies turned the same embed-and-ship posture into an Australia → Silicon Valley talent export for YC-class product work. This article separates the Palantir blueprint, Lyra's forward-deploy model, the current AI-FDE market (including the viral "$1M/year" framing), the audit → evals → deployment loop from Greg Isenberg's conversation with Vas of Varick Agents, and what the same job looks like when the customer is a Solana protocol, fintech, or ecosystem team.

Click to zoom

What an FDE is (plain definition)

An FDE is a strong engineer who works forward — next to the customer or business owner — not only behind a product roadmap. They own more of the path than a pure solutions engineer or pure product engineer:

• Discovery — watch how work really happens (exceptions, spreadsheets, tribal knowledge)
• Scoping — pick a valuable, bounded system edge
• Build — integrations, app logic, model calls, review gates
• Evals — turn non-determinism into evidence
• Deploy + adopt — production, monitoring, training, iteration

OpenAI's public FDE listings describe roughly that arc: discovery through production rollout, then success measured by adoption, workflow impact, and eval-driven feedback. If your "FDE" only demos and never ships, you hired a salesperson with a laptop.

Palantir: the original blueprint

Palantir popularized Forward Deployed Software Engineers (FDSEs): embed with hard customers (defense, intelligence, later commercial verticals), learn the domain fast, make architecture decisions, and carry custom applications from idea to deployment. Travel, stakeholder work, and code in the same week are features of the role, not bugs.

The conceptual artifact worth stealing is the Ontology: an operational layer that maps data and models onto real-world objects, relationships, actions, and governance. In plain English: model how the company actually works, not only the tables it stores. You do not need Foundry to use the mental model. For any FDE project, map four things before you pick a model:

1. Data     — facts required; who may access them
2. Logic    — rules vs judgment vs model vs human
3. Action   — what changes in real systems after the decision
4. Security — who can initiate, approve, inspect, reverse, audit

That is a better starting question than "where can we add an agent?"

Lyra: forward deployment as an Australian export

Palantir is the origin story. Lyra Technologies is one of the cleanest modern examples of the same posture productized as a company — not selling a platform ontology, but forward-deploying engineers (and designers) into other people's startups.

Lyra's pitch is explicit: since 2022 they have partnered with Silicon Valley startups to design, build, and ship product from studios in Sydney and Melbourne (with SF presence). Public numbers on the site: 75+ client engagements, 15+ YC companies, 70+ headcount. Case studies read like classic FDE outcomes — embedded shipping while a founding team is still hiring (e.g. Lucent), agent-heavy product work through seed (Hobbes), CLI/CI and frontend velocity for YC infra companies (Thunder Compute). Client language repeats the same refrain: extension of the core team, not a vendor ticket queue.

Their careers path even names the ladder in FDE vocabulary: ship real product next to strong peers, then step into a full FDE role embedded inside a Silicon Valley startup as part of that company's team. AI tooling reimbursement is table stakes; expected use of models is part of craft, not a side experiment. That is the market making the Isenberg/Vas thesis concrete in APAC: the scarce unit is a high-judgment engineer who can land inside someone else's codebase and produce outcomes.

Why Lyra matters for this article's readers:

• Proof the model travels. You do not need to work at Palantir to practice forward deployment; networks like Lyra (and boutique AI implementers such as Varick) are institutionalizing it.
• Australia ↔ SF arbitrage. Talent and timezone leverage for YC-class startups that need product velocity without waiting on a full local hire plan.
• Career path clarity. Embed → ship → earn trust → deeper ownership is the same loop whether your badge says Delta, FDE, or "Lyra eng on the client Slack."
• Solana / crypto adjacent. The same muscle applies when a protocol or fintech needs an embedded engineer for integrations, agent rails, or GTM-critical tooling — not a six-month RFP for a slide deck.

(Note: do not confuse Lyra Technologies AU with Lyra Technology Group, a separate US multi-business services family that also posts "Forward Deployment AI Engineer" roles for portfolio companies. Same three-letter idea, different org.)

Why the role exploded in AI

Model capability went up and access went wide. Implementation stayed specific. MIT NANDA's widely cited 2025 GenAI work is often reduced to a soundbite ("95% of pilots fail"); the careful reading is closer to: most orgs saw no measurable return in the study window, and few task-specific tools reached production — with methodology caveats. The operational lesson still holds: workflow fit and measurement beat demos.

Hiring markets responded. Job boards and surveys through 2025–2026 show large growth in FDE / applied-AI deployment titles across frontier labs, applied AI startups, and enterprise AI teams. Comp is tiered hard:

• Classic Palantir-style FDSE bands — public US listings often land roughly mid–high six figures base depending on city and level (example NY FDSE bands have published in the ~$135–200K salary range plus equity/incentives).
• Frontier-lab FDE bands — OpenAI-class public base ranges have sat roughly ~$162–280K plus equity; total package can be much higher at senior levels when equity hits.
• The "$1M FDE" headline — possible at the extreme (seniority + equity + scarcity + commercial impact), not a median. Treat podcast ceilings as marketing; build the skill stack before optimizing the number.

The Isenberg / Vas masterclass: clear definition of the work

On Greg Isenberg's Startup Ideas Podcast (episode framing: FDE: The $1M/Year AI Job Explained, Jul 2026), Vas of Varick Agents walks the role without mystical language. Premise: intelligence is commoditizing; deployment is the edge. He traces FDEs to Palantir-style embedding and customizing to the client's ontology, then argues the scarce combo is engineering judgment + consulting-grade communication — art plus science.

Episode spine (timestamps from the published show notes):

02:03  What is an FDE
04:09  How Palantir popularized FDEs
06:16  Deciding where intelligence belongs
11:26  What FDEs earn (ceiling vs expectation)
14:59  Two kinds of judgment: communication + engineering
20:40  Audit → evaluation → deployment
27:36  Audit: find the workflow worth rebuilding
31:47  Evals: turn non-determinism into evidence
32:57  Deployment: build on systems people already use
38:59  30-day plan: build, harden, measure, defend

Companion material from the episode: FDE Blueprint and Varick's FDE in 30 Days. Use them as training routes, not as a guarantee of title or pay.

Where intelligence belongs

Vas's sharpest practical cut: not every step deserves an LLM.

• Deterministic code — rules, arithmetic, authz, schemas, compliance gates
• Model — ambiguous language, unstructured docs, flexible classification, synthesis, tool planning
• Human — high-consequence, novel, regulated, irreversible decisions
• Hybrid (most real systems) — model proposes → code validates → policy checks → human approves → system executes → logs capture everything

Use models where uncertainty creates value; use software where certainty is available; put humans where the cost of a wrong action exceeds the speed of automation.

The core loop: audit → evals → deployment

1. Audit the real workflow

Interview operators, not only executives. Walk the last real case including the awkward spreadsheet and the exception path. Score opportunities on impact, repeatability, data readiness, integration effort, failure cost, ownership, and measurability. First project should be bounded and owned — not the flashiest agent fantasy.

2. Evals: evidence over vibes

Evals are tests that check system outcomes against criteria you specify. Unit of measure is usually the complete workflow outcome, not a polite sentence from the model. Build a set with normal cases, hard cases, known failures, adversarial inputs, and required escalations. Track task success, unauthorized actions (target zero), escalation quality, latency, cost per success, adoption, and the business metric the sponsor already funds.

3. Deploy into existing systems

Build on Salesforce / SAP / NetSuite / Slack / the internal API — whatever operators already live in. Autonomy ladder:

Offline replay  →  Shadow mode  →  Assist (human approves)
→  Bounded auto (low-risk only)  →  Expanded autonomy (earned by evidence)

"The model is better now" is not a release criterion. Pass rates, zero unauthorized actions, cost ceilings, and shadow weeks are.

30-day portfolio plan (do the job before the title)

Week 1  Complete the loop
        One real workflow, min tools, human checkpoint, audit trail
        Proof: demo + map + definition of done

Week 2  Harden
        Schemas, retries, timeouts, idempotency, permissions, exceptions
        Proof: failure table + recovery demo + runbook

Week 3  Measure
        Golden set, failure taxonomy, cost/latency, baseline compare
        Proof: eval report + business-impact estimate

Week 4  Defend
        Architecture narrative, security boundaries, rollout, economics
        Proof: 5-min demo + 1-page business case + risk register

Thirty days will not make you a principal FDE. It can produce a portfolio artifact that proves you understand delivery beyond prompting — especially if you include a deliberate failure demo (broken API, prompt injection, high-risk case) showing stop, escalate, log, recover.

Two kinds of judgment

Weakness on either side sinks deployments:

• Engineering judgment — system boundary, deterministic vs probabilistic, schemas, retries, evals, observability under load
• Communication judgment — how work really happens, translating tradeoffs into risk/cost/time, setting expectations about uncertainty, earning trust so users report failures instead of abandoning the tool

Many engineers avoid messy stakeholders. Many consultants cannot repair the system they recommend. The FDE is the overlap — which is why the market pays for it.

Solana needs FDEs too

Crypto rebranded the same job for years: solutions eng, integrations eng, ecosystem eng, technical partnerships, on-site hackathon mentors, DevRel who actually ships. The FDE label is useful because it insists on production outcomes, not only content or lead-gen.

High-leverage Solana FDE surfaces right now:

• Protocol integrations — embed with a partner, map their settlement/risk workflow, ship a working integration with monitoring (see our bank/stablecoin and payments coverage)
• Agentic payments and x402-style rails — bounded autonomy, evals on tool use, human gates on value movement
• Enterprise or fintech on-ramps — the audit/eval/deploy loop against compliance and ops, not a weekend dApp
• Builder success that is measured — time to first mainnet tx, retention on docs/tools, support deflection — DevRel with an FDE delivery spine

The ontology question maps cleanly: accounts, programs, mints, vaults, authorities, and who can sign what are objects, actions, and security — not a slide deck of features.

Who to hire: Sal Samani (FDE-shaped Solana profile)

If you are staffing an FDE or FDE-adjacent role for Solana, AI agents, ecosystem delivery, or Australia/APAC embedding — the same corridor Lyra proves works for YC product velocity — consider Sal Samani (linkedin.com/in/metasal, @metasal, metasal.xyz).

Mapped against the FDE skill stack from this article and the podcast:

• Embed + community ontology — Head of Australia for Solana ANZ; operates Superteam AU context; runs meetups and builder networks that surface real workflows, not survey fiction.
• Ship systems, not only slides — 40+ tools, apps, and experiments (passkeys, payments, explorers, agent experiments including Clawbook / Colosseum honourable mention); founder of Milysec (Solana / AI venture studio).
• Delivery surface for developers — creator of devrels.xyz (directory, technical articles, community ops): the DevRel side of "adoption + measurement" that FDE job specs keep asking for.
• Communication + engineering judgment — background as a software engineer who also does stakeholder and ecosystem work; bilingual in builder rooms and executive-style narrative.

That is the art-plus-science profile Vas describes — applied to Solana and AI tooling rather than a single enterprise ERP. For hiring managers: the conversation is open via LinkedIn; for candidates studying FDE, his public shipping record is a portfolio pattern to copy (many small production systems + community embedding + measured developer platforms).

Titles that hide the same job

Search beyond the exact three letters. Related labels: forward deployed software engineer, applied AI engineer, AI solutions architect, implementation engineer, field engineer, technical deployment lead, integrations engineer, ecosystem engineer. Read the description for embed + build + measure + own production. If it is only demos and RFPs, it is sales engineering — valuable, different.

Bottom line

Palantir proved that hard problems yield when engineers sit inside the customer's operational reality. Lyra shows the same posture can be a standing talent export from Australia into Silicon Valley product teams. AI made the structure urgent because models are abundant and reliable workflows are not. Vas's contribution on the Isenberg show is a usable loop — audit, evals, deploy, defend — and a 30-day path to prove you can run it.

Do not chase the million-dollar headline. Chase one repeated workflow, a golden eval set, a human gate on high-risk actions, and a sponsor who can define done. That is the FDE job. On Solana, the workflows are programs, payments, agents, and builder success — and the people who can embed and ship are still scarce.

Resources

• FDE: The $1M/Year AI Job Explained — Greg Isenberg × Vas (Varick Agents)
• Varick Agents · FDE in 30 Days · @vasuman
• Palantir Ontology overview · FDSE job listings on Palantir careers / Lever
• Lyra Technologies — AU forward-deployed eng/design into SV / YC startups · careers (FDE path) · LinkedIn
• OpenAI careers — FDE postings (discovery → rollout → evals)
• OpenAI evals guide · NIST AI RMF · OWASP AI Agent Security
• Hire / contact: Sal Samani on LinkedIn
• Related on devrels.xyz: Solana AI agents 2026, agentic payments landscape, x402 vs agent pay