Home › Forward deployed engineer roadmap

Forward deployed engineering · start here · one page

The Forward Deployed Engineer roadmap for data and AI

A forward deployed engineer sits with a customer, finds the real problem and ships a working, evaluated, costed system in the customer's own environment, then hands it over. For data and AI that means being part data engineer, part AI engineer, part software engineer, and part technical anchor, business analyst and delivery lead. This page is the map: our recommended roadmap, and a way into every guide on the site.

The role on one picture

Four kinds of skill, one outcome

The job is the combination, not any one part. A data platform with no discovery solves the wrong problem; a clever agent on poor data fails in use; a good build the customer cannot run is not finished.

DataData engineeringPlatform, cost, securityAIAI engineeringevals, guardrails, harnessEngineering craftSoftware engineeringTechnical anchorCustomerBusiness analysisAgile and deliveryCustomer craftForward Deployed Engineer, Data and AIships a working, evaluated, costed system the customer's team can own

Eight tracks in four clusters. Each track has a card below with what to be able to do and where to learn it.

Our recommendation

The roadmap we would follow, and why

Do not learn the tracks one after another. Learn them in parallel at rising depth, against a real engagement, with one piece of evidence at the end of each phase. Seven principles sit behind that.

PrincipleWhyWhat to do
Deep in data, working across the restA T shape. Data engineering is the deep stroke because AI systems are only as good as the data under them. The other tracks need working skill, not mastery.Pick data as your depth; reach 'build' level in every other track before aiming for 'own'.
Parallel tracks, not a queueA forward deployed engineer is never only technical. If customer craft waits until year two, you build the wrong thing very well.Touch discovery, estimation and writing from week one, in small doses.
Learn against one real engagementCourses teach vocabulary; a capstone teaches decisions. Choose one realistic problem and let it pull the learning.Keep one capstone repository and one decision log for the whole roadmap.
Deterministic firstAI engineering sits on top of data and software engineering; it does not replace them. Use code where code works and a model only where judgement on language is needed.Before adding a model, write what plain code could do and what it cannot.
Evidence over demosDecision makers approve what they can check. Every phase should end with an artifact someone else can read.End each phase with a one-page evidence note: what you built, how you measured it, what it costs.
One platform deep, others by ideaTools change; ideas last. Go deep on one cloud and one data platform and learn the others as concepts.Choose your platform from the learning paths, and map every concept to it once.
Teach it backWriting an idea in plain words shows exactly where you do not understand it, and is how you earn trust with customers.Write one short note per week. Publish the good ones.

When to learn what

Eight tracks across four phases

The grid shows how deep to go in each phase: taste (know it exists and try it), build (do it with help), own (do it alone and teach it). Click a track name to jump to its card.

1 Foundation2 Build3 Deliver4 LeadData engineeringownownbuildbuildPlatform, cost, securitytastebuildownownAI engineeringtasteownownownSoftware engineeringownbuildbuildbuildTechnical anchortastebuildownownBusiness analysistastebuildownownAgile and deliverytastebuildownownCustomer crafttastebuildownown

Depth by phase. Foundation leans on data and software; Build adds AI; Deliver and Lead are customer-facing. Time spans are rules of thumb.

1. Foundation

about months 0 to 3

Be a safe pair of hands on data and code. Data engineering and software engineering come first, because everything later stands on them. Touch every other track lightly.

You are ready when
  • You have built a small pipeline from raw to gold, with tests, in git.
  • You can explain why a table is laid out the way it is.
  • You can write a customer problem in one paragraph.

Leave behind: One small pipeline repository and a one-page problem statement.

2. Build

about months 3 to 6

Add AI engineering and shape the platform. Build an agent with tools and a human approval step, and evaluate it on agreed cases. Learn what it costs per run. Start discovery and estimation practice.

You are ready when
  • You have an agent with an evaluation set and a cost per case.
  • You can say what the model will not do, and why.
  • You can read a slow or expensive job and name the bucket.

Leave behind: An agent, its evaluation report and its cost page.

3. Deliver

about months 6 to 12

Run a real engagement, from discovery to hand-over. Discovery, prototype in days, deploy into the customer's cloud, show evidence, write the runbook. This is where the customer-facing tracks become the main work.

You are ready when
  • It runs in their account, not yours.
  • You can show quality and cost on one page.
  • Their engineers run it without you.

Leave behind: An engagement case study with the evidence page and runbook.

4. Lead

year 2 onward

Make the second customer faster than the first. Reusable assets with a port and an adapter per customer, architecture ownership, mentoring, and saying no with evidence.

You are ready when
  • A second customer reuses your framework with a new configuration.
  • Others rely on your reviews and templates.
  • You choose which work not to do.

Leave behind: A reusable framework and a design review you led.

Months are a guide for someone working on this steadily. If you already work in data engineering, Foundation compresses a lot. The checkpoints matter more than the calendar.

Where are you now

Rate yourself on each track

Be honest: pick the highest level where you could do all of it, with no help. The page shows your weakest track and links to it. Your choices stay in this browser only.

TrackYour levelOpen
Data engineeringcard
Platform, cost, security and governancecard
AI engineeringcard
Software engineeringcard
Technical anchor and design decisionscard
Business analysis and discoverycard
Agile, delivery and estimationcard
Customer craft and judgementcard
Pick a level for each track to see where to focus next.

The eight tracks

What good looks like, and where to go next

Each card says what you should be able to do at crawl, walk and run level, and links to the guides, simulators and learning paths for it. Find your level on each card, then open the next link.

Data engineering

Moving, shaping and trusting data so every use case has something solid to stand on. This is the base the AI work sits on.

CrawlSQL and Python fluency. Read a source system. Build raw to gold on a small dataset, with tests. Know why file layout matters.WalkIncremental loads, late data and history (SCD). Quality checks. Read a slow job and say where the time goes. One platform in depth.RunDesign a customer's platform: layers, contracts, environments, cost model. Review someone's pipeline in ten minutes. Teach it.
Data engineering hubFundamentals: six mind-map guidesData systems, end to endMedallion architectureShift left: fail fast on data codeLearning path: data engineering

Platform, cost, security and governance

The cross-cutting concerns that decide whether a system is allowed to run and affordable to keep running.

CrawlKnow what a lakehouse, a catalog and a cloud account are. Basic identity and least privilege. Read one line of a cloud bill.WalkRight-size and schedule compute. Set up environments and CI. Apply masking and row rules. Read a query profile.RunRun a cost and security review of a customer platform. Set guardrails and policies. Estimate run cost before building.
Performance fundamentalsCost fundamentalsSecurity fundamentalsStorage fundamentalsData platform architectureMind map: every concernAI governance

AI engineering

Putting a model to work inside a system that is tested, bounded and costed. Deterministic first, a model only where judgement on language is needed.

CrawlCall a model. Structured output and tool calling. Know tokens, context and hallucination. Apply the deterministic-first rule.WalkBuild an agent loop with tools and human approval. Evaluate on agreed cases. Add guardrails and tracing. Know the cost per case.RunDesign a harness many customers can reuse. Set evaluation gates. Choose a model per job. Run it with governance.
Learning path: AI engineeringThe AI universeAgent Harness: intro courseWhich AI model for which jobEval HarnessAgent guardrailsAgent frameworks: when to use whichClaude API simulator

Software engineering

The craft that makes the rest shippable: version control, tests, interfaces, deployment and operating what you build.

CrawlGit, branches and code review. Tests. APIs. Packaging. Reading other people's code without fear.WalkCI/CD, infrastructure as code, containers, observability. Design for failure and retries.RunPorts and adapters, templates and golden paths, decision records. Swap a vendor by writing one adapter.
Learning path: software engineeringKey concepts: softwareLearning path: LLM engineering

Technical anchor and design decisions

The person who keeps the technical truth of a project: asks the hard questions early, writes decisions down, reviews against checks.

CrawlUnderstand the role. Write one decision with options and a recommendation. Know what to ask in technical discovery.WalkRun a technical discovery. Review a data project against stage checks. Raise risk while it is cheap.RunOwn architecture across a programme. Mentor. Say no, with evidence.
The technical anchorTechnical discoveryProject review checksCritical thinking in data projects

Business analysis and discovery

Finding the real problem in the customer's words, agreeing how success is measured, and writing down what the system will not do.

CrawlInterview the people who feel the pain. Write the problem in one paragraph they agree with. Learn the source-system questions.WalkMap data and owners. Write requirements that say what is out of scope. Reconcile numbers with the business.RunRun discovery across several workflows. Prioritise by value and effort. Build the business case.
Business analysisDiscovery: the problem in their wordsAgree the measure of successScope by what it will not doData map and ownersSource system questionnaireWhere data engineering time goes

Agile, delivery and estimation

Slicing work thin, estimating honestly, showing working things often, and keeping the customer's expectations and yours the same.

CrawlSprints, backlog, definition of done. Slice work thin. A short weekly update.WalkEstimate with the unknown source in mind. Manage approvals and dependencies. Demo something working every cycle.RunPlan multi-team delivery. Manage scope and expectations. Avoid the classic data-project mistakes.
Project managementEstimation mistakes in data projectsTop project management mistakes

Customer craft and judgement

Writing, explaining, showing evidence, handing over. The skill that turns a good build into an engagement the customer keeps.

CrawlOne-page writing. Ask five questions of any claim. Explain without jargon.WalkEvidence pages instead of demos. Steering meetings. Handle disagreement. Write a runbook.RunHand over so you are no longer needed. Build reusable assets. Earn trust over years, not weeks.
Evidence, not demosHand-over and runbooksOne framework, many customersMindsetFirst 30 daysPrototype vs product

The engagement

One loop that ties the tracks together

Every engagement runs the same nine steps. Each step is a concept page, and each uses a few of the tracks above.

Discoverdiscovery1Measuresuccess2Scopescope-out3Map datadata map4Prototypein days5Proveevidence6Deploytheir cloud7Hand overrunbook8Reuseassets9the next engagement starts from the assets of the last

Click a box to read the concept. The last step feeds the first step of the next customer.

StepWhat happensMostly uses
DiscoverTalk to the people who feel the pain and watch the work as it is done today.Business analysis
MeasureAgree one or two numbers that mean success, and measure today's value.Business analysis
ScopeWrite what the system will not do. Code for the repeatable parts, a model where language needs judgement.Technical anchor, AI engineering
Map dataWhere the data lives, its shape, who can say yes. Most delay is access, not code.Data engineering
PrototypeA working demo on their real data in days, not slides.Software, data and AI engineering
ProveEvaluation on agreed cases, errors listed, cost per case, traces behind examples.AI engineering, customer craft
DeployInto their cloud, under their identity, network and approval rules.Software, platform, security
Hand overRunbook, training, shadowing, clear ownership.Customer craft
ReuseKeep the framework; bind it to each customer with an adapter and configuration.Software engineering, technical anchor

Practise on something real

Capstone projects

Choose one and keep a decision log. Real constraints teach what courses cannot.

CapstoneWhat it teaches
Claims first notice of loss on AzureAn agent that reads claim emails and fills the form, deployed in a customer-style Azure setup. Good for the Deliver phase.
Clinical data review agentA data layer plus an agent that finds issues in clinical trial data, evaluated honestly. Good for the Build phase.
Eval Harness frameworkThe evaluation piece on its own: stores, metrics and runs. Good for learning what 'evidence' means.
Your own real workflowPick a workflow you or a colleague repeats weekly, write the one-page problem, and run the loop above on it.

Your week

A weekly rhythm that keeps all tracks warm

Share of your weekWhatWhy
About 40%Build on the capstoneDecisions are learnt by making them. Keep a decision log.
About 20%Learn one track on purposeRotate the track each fortnight so none goes cold.
About 20%Customer craftInterview someone, write a problem statement, give a short demo, ask for feedback.
About 10%Read other people's systemsCode review, a postmortem, an architecture. It is the cheapest way to see mistakes.
About 10%Teach it backOne short note a week, in plain words.

What to leave for later

Temptations, with a verdict

TemptationVerdictWhy
Learning every toolDropTools churn. Learn one platform deep and the rest as concepts.
Collecting certifications before buildingDelayBuild something first; take a certification when it confirms what you already do.
Fine-tuning models earlyDelayMost first problems are solved by data, prompts, tools and evaluation. Fine-tune only when an evaluation shows the need.
Hopping between agent frameworksDropLearn the loop, tools, state and approval ideas once; frameworks are wrappers.
Deep machine learning theoryDelayUseful if your engagements need modelling; not a prerequisite for most forward deployed work.
Slide-first demosDropShow a working thing on their data, then the evidence page.
Streaming for everythingDelayReal time costs more. Use it only where the business acts within seconds.

Certifications

Where they fit

Build first, certify to confirm. A sensible order: a platform or cloud certification near the end of Build, an AI certification once you have an agent with an evaluation, and an architect-level certification after your first Deliver. Options by platform are on the certifications page, and the forward deployed learning path lists them per cloud.

FAQ

Questions people ask

What is a forward deployed engineer, in this roadmap?

An engineer who sits with a customer, finds the real problem and ships a working, evaluated, costed system in the customer's own environment, then hands it over. Here it covers data, AI, software, technical anchor, business analysis and agile delivery.

Do I need to be expert in all eight tracks?

No. Be deep in one, usually data engineering, and reach working level in the others. Own the customer-facing tracks by the Deliver phase because that is where engagements are won or lost.

How long does this take?

The spans are rules of thumb for someone working on it steadily. If you already work in data, the Foundation phase compresses a lot. If you are starting out, expect it to stretch. The checkpoints matter more than the months.

Data first or AI first?

Data first, in small doses of AI from the start. AI built on weak data and weak software is the usual cause of agents that demo well and fail in use.

Where do certifications fit?

After you have built something. Use them to confirm a level, not to reach it. The certifications page lists the options by platform.

Is this the only valid roadmap?

No. It is our opinion, shaped by the pages on this site. Use the tracks as a menu and change the order to fit the work in front of you.

Related

All the starting points

Data engineeringFundamentalsAI engineering pathSoftware engineering pathForward deployed learning pathTechnical anchorHow we workMindsetCertificationsProjects to joinKey concepts, A to Z

This roadmap is our opinion, built from the pages on this site. Time spans are rules of thumb, not promises.