A written version of the interactive roadmap above — every station, what you'll learn, and a small thing to build — laid out for reading, reference and search.
The Craft Start here
F1. The FDE Role
Beginner · 40 min
Forward-deployed means embedded: you sit with the customer, own an outcome instead of a feature, and carry the product into a messy real-world environment. Learn where the role came from, why every AI lab now hires for it, and the tension you will manage forever — building for one customer while feeding a product that serves many.
Skills: What forward-deployed means · Outcomes over features · Product vs consulting tension · Where FDEs fit in AI labs
Build it: Take a real product you know and write the one-page brief for deploying it inside a specific company: the outcome, the blockers, the first two weeks.
F2. Rapid Prototyping
Beginner · 55 min
The FDE superpower is the 48-hour demo: something real, on the customer’s actual data, days after the kickoff. Learn to scaffold fast with coding agents, fake the parts that do not matter yet, and get real data flowing early — because a mediocre demo on their data beats a perfect demo on yours.
Skills: 48-hour demo muscle · Agentic coding speed · Fake the right parts · Real customer data early
Build it: Scope a demo you could ship in two days for a claims-processing team: what is real, what is mocked, and what data you need from them on day one.
F3. The LLM Toolkit
Intermediate · 60 min
Your raw material is the model. Learn the working toolkit — prompting that survives contact with users, structured outputs your code can parse, picking the right model for the job, and the token math that decides what a workflow costs at the customer’s volume.
Skills: Production prompting · Structured outputs · Model selection · Cost-per-workflow math
Build it: A customer wants 50K support tickets classified daily. Pick the model, design the output schema, and estimate the monthly bill.
F4. Enterprise RAG
Intermediate · 65 min
The default FDE deliverable: the customer’s knowledge, answerable. Learn chunking, retrieval and reranking, grounding with citations, and access control — because in an enterprise, who may see which answer matters as much as the answer.
Skills: Chunking & retrieval · Grounded, cited answers · Access-controlled RAG · Retrieval quality tuning
Build it: Design RAG over a company wiki where sales must not see HR docs. Where does permission filtering happen — index, query, or prompt?
F5. Agents & Workflows
Intermediate · 60 min
Past Q&A lies automation: agents that file the ticket, update the record, draft the response. Learn tool calling, the plan-act-observe loop, and the judgment call FDEs make weekly — when the customer needs an agent, and when a boring script is the better product.
Skills: Tool calling · Workflow automation · Agent vs script judgment · Bounded autonomy
Build it: A customer wants "an agent for invoice processing". Sketch the workflow, then mark which steps deserve a model and which deserve code.
F6. Integration Glue
Intermediate · 55 min
Enterprise value lives inside Salesforce, SAP, SharePoint and twenty internal tools. Learn the glue: REST and webhook plumbing, MCP for standardized tool access, auth flows (SSO, OAuth, service accounts), and reading undocumented legacy APIs without crying.
Skills: API & webhook plumbing · MCP integrations · SSO / OAuth / service auth · Legacy system archaeology
Build it: Map the integration plan for an AI assistant that reads CRM records and writes follow-up tasks: systems, auth, and failure points.
The Field On site
T1. Discovery & Scoping
Intermediate · 55 min
The stated problem is rarely the real one. Learn to run discovery — shadow the workflow, find where hours actually go, define success as a number before writing code — and to scope ruthlessly: the pilot that ships beats the platform that does not.
Skills: Workflow shadowing · Success metrics upfront · Ruthless pilot scoping · Saying no to magic
Build it: A COO says "we want AI for operations". Write the five discovery questions that turn that sentence into a scoped, measurable pilot.
T2. Customer Data Wrangling
Intermediate · 65 min
The data is always worse than promised: scanned PDFs, SharePoint labyrinths, CSVs exported from a system nobody remembers. Learn document extraction, cleaning pipelines, and triage — which data problems you fix, which you work around, and which kill the use case honestly and early.
Skills: Document AI & OCR · Cleaning pipelines · Data triage · Kill criteria
Build it: The customer hands you 10K scanned invoices and an Excel of vendor codes. Design the pipeline from mess to queryable — and name the step most likely to fail.
T3. Deploy Anywhere
Advanced · 65 min
Your beautiful stack meets their VPC, their on-prem cluster, or an air-gapped room. Learn to package for hostile territory — containers and Kubernetes, egress rules, model access constraints (their cloud, your API, or local weights) — and to debug through a bastion host with someone watching.
Skills: Containers & Kubernetes · VPC / on-prem / air-gapped · Model access constraints · Remote debugging
Build it: The customer allows no data egress. Redesign your architecture: where does the model run, and what breaks compared to your SaaS setup?
T4. Security & Compliance
Advanced · 60 min
Before production comes the gauntlet: the security review, the DPA, the compliance checklist. Learn to speak SOC 2, GDPR and HIPAA at working depth, handle PII (redaction, retention, residency), and wire guardrails so the AI cannot be talked out of the rules.
Skills: Security review survival · PII handling & residency · SOC 2 / GDPR / HIPAA literacy · Guardrails as policy
Build it: Prepare the one-page answer to "where does our data go?" for a RAG deployment — every hop, every store, every retention clock.
T5. Evals for Client Work
Advanced · 60 min
The customer will judge the system on examples you never saw. Learn to build the golden set from their real cases, agree acceptance criteria before launch, score with LLM-as-judge where labels are scarce, and rerun the suite before every change — so "did it get worse?" is a lookup, not a fight.
Skills: Golden sets from real data · Agreed acceptance criteria · LLM-as-judge scoring · Regression before release
Build it: Define the eval for a contract-review assistant: 30 documents, what is measured, and the number below which you do not ship.
T6. The Demo
Intermediate · 50 min
Deals move on demos. Learn the craft: their data on screen, one rehearsed wow moment, graceful handling when the model fumbles live, and the discipline to show real limits — because a demo that overpromises writes a check the deployment cannot cash.
Skills: Demo narrative design · Their data, live · Failure choreography · Honest wow
Build it: Script a 10-minute demo of a RAG assistant for a skeptical buyer: the opening question, the wow beat, and the planned recovery when an answer is weak.
The Outcome Land & expand
P1. Stakeholder Craft
Advanced · 55 min
Deployments die politically before they die technically. Learn to map champions and blockers, translate model behaviour into business language, and manage the 80% problem — the demo amazed everyone, now you must explain what the other 20% costs and who handles it.
Skills: Champion & blocker mapping · Translating model limits · Expectation management · Executive communication
Build it: Draft the three-paragraph email that resets expectations after a flashy demo: what ships now, what needs data, what stays human.
P2. Pilot → Production
Advanced · 65 min
The prototype won the pilot; now it must survive Monday morning. Learn the hardening pass — auth, logging, tracing, SLOs, on-call handoff — and the discipline of turning demo shortcuts into either real engineering or documented debt.
Skills: Hardening checklist · Tracing & logging · SLOs & alerts · On-call handoff
Build it: List the ten things that must change before your demo RAG app serves 500 employees — and which three you refuse to launch without.
P3. Reliability & Incidents
Advanced · 60 min
It broke inside their network, and you cannot SSH in. Learn safe rollout machinery — feature flags, staged releases, kill switches — plus remote debugging etiquette and the incident write-up that turns a bad Tuesday into renewed trust.
Skills: Feature flags & staged rollout · Kill switches · Remote incident debugging · Blameless write-ups
Build it: Design the rollout plan for a risky model upgrade at a customer: stages, the metric watched at each, and the exact rollback trigger.
P4. Proving Value
Advanced · 55 min
Renewals are bought with numbers, not vibes. Learn to instrument the metrics the buyer cares about — hours saved, deflection rate, accuracy versus the old process — baseline them before launch, and ship the dashboard the executive actually opens.
Skills: ROI metric design · Baselines before launch · Deflection & time-saved · The exec dashboard
Build it: For a support-triage deployment, define the three numbers on the renewal slide and exactly how each is measured.
P5. Scale & Repeatability
Advanced · 60 min
One heroic deployment is a story; ten repeatable ones are a business. Learn to turn bespoke work into playbooks and templates, keep customer variants as configuration instead of forks, and feed field learnings back to the product team as their sharpest signal.
Skills: Playbooks & templates · Config over forks · Field-to-product feedback · Multi-customer serving
Build it: After three similar deployments, extract the template: what is identical, what is config, and what was never worth customizing.
P6. The Career
Intermediate · 45 min
FDE is a launchpad: the rare role that compounds engineering, product sense and customer judgment at once. Learn the paths out — solutions architecture, product, founding — why AI labs pay a premium for it in 2026, and how to keep your engineering edge sharp while living in the field.
Skills: FDE → SA / product / founder · Compounding skill set · Staying technical in the field · Reading the market
Build it: Write your six-month FDE development plan: one craft skill, one field skill, one outcome skill — and the deployment that proves each.