Contents of Become a Forward Deployed Engineer: Ship AI for Real Customers
The whole course told through one customer, what you should already know, and an AWS account and a model key set up safely before the first lab. The course is planned as 63 lessons: the opening lessons and 8 parts. 8 lessons are ready to study today, about 6 hours of learning, with 4 hands-on labs, 3 of them in your own AWS account, and more than 150 diagrams and real screenshots. New lessons are added as they are finished, and you get every one. The free lessons open for everyone; the rest open when you buy the course.
See the early-access price and what the course covers
Your progress0 of 7 checks answered right
It fills only when a prediction matches the measured result. Reading a page does not move it. It is kept in this browser.
Start here
The opening lessons
The whole course told through one customer, what you should already know, and an AWS account and a model key set up safely before the first lab.
2 of 3 lessons published
- Introduction: what this course buildsThe whole course told through one customer, Bramfield Mutual, from the first call to the day its claims team uses the system without you. Free
- Before you startThe Python, SQL, terminal and Git this course assumes, the tools to install, and a five-minute self-check that ends with your AWS account answering. Free
- Your AWS account and model key, set up safelyA user that is not root and holds only the permissions the labs need, single sign-on instead of long-lived keys, a spending limit on the model key, and a check that the budget alert from "Before you start" is working.Coming
Part 1
The job and the customer
What the job really is, read from the companies that hire for it, how it differs from the roles around it, and the customer whose systems you inherit for the rest of the course.
6 lessons, all published
- What a forward deployed engineer doesOne customer, many capabilities: how Palantir, OpenAI, Salesforce and Ramp describe the work in their own words, and what a real week looks like. Free
- How the job differs from solutions engineer, consultant and product engineerWho writes production code, who owns the outcome and who leaves after the slides, read from real job posts. Locked
- Meet Bramfield Mutual, the customer in this courseA fictional Pennsylvania insurer with a claims backlog, an old claims database, nightly file drops and strict rules on personal data; you build its systems on AWS and read the brief. Free
- How a customer engagement runsDiscovery, pilot, production, handover and expansion, what each phase must produce, and the pod of two engineers and one strategist. Locked
- Your first day inside systems you did not buildMapping Bramfield's servers, databases, file drops and services with read-only access, and drawing the system map you will use all course. Locked
- Working notes, a decision log and a daily updateThe three documents that let anyone else pick up your account, kept in Git from the first day. Locked
Part 2
Discovery and scoping
The first meeting, the people who matter on the account, a baseline measured on the customer's own data, the security questions for week one, and the scope of work for the pilot.
7 lessons, coming
- The first meeting: "we want AI for claims"The questions to ask in the first hour, how to hear the real problem under the request, and the note you send afterwards.Coming
- Who matters on the accountThe head of claims, the adjusters, IT, security, compliance, vendor risk and the person who signs, and what each of them needs from you.Coming
- Turning a wish into a number you can measureMeasuring today's days-to-settle, backlog and supervisor referrals from Bramfield's own claims database, and what each day of delay costs in dollars, so the pilot has a baseline.Coming
- Is the data really there?Profiling the claims tables, the file drop and the documents for gaps, duplicates and owners before you promise anything.Coming
- Security and compliance discoveryThe state insurance data-security law, California privacy law, whether HIPAA applies, the SOC 2 questionnaire, and the request to keep data in US regions: what to ask the security team in week one.Coming
- When the answer is not an LLMThree Bramfield requests where a rule, a query or a better form beats a model, measured on the same 200 claims, and how to say so politely.Coming
- The scope of work and the pilot planWriting the pilot scope, the success measures, what is out of scope, the timeline, the risk register and the pilot's cost in dollars, using the numbers you measured.Coming
Part 3
Getting into the customer's systems
The integration work an engineer spends whole weeks on: least-privilege access, the customer's single sign-on, an old SOAP service, a nightly file drop, incremental loads, safe write-back, private networking, the corporate proxy and secrets.
9 lessons, coming
- Access the right way: a scoped account, not global adminAsking for and testing least-privilege access to Bramfield's database, files and AWS account, and refusing the "admin to save time" offer.Coming
- Logging in through the customer's identity providerSigning Bramfield's staff into your app with OIDC and SAML through Keycloak, and checking what happens the day someone leaves.Coming
- Calling an old SOAP API: paging and rate limitsReading 20,000 policies from Bramfield's SOAP policy service without overloading it, with retries, backoff and a resumable cursor.Coming
- The nightly file dropPicking up CSV exports from Bramfield's SFTP server, with two date formats, amounts written as "$1,200.50" and "1200.5 USD", and duplicate rows.Coming
- Loading only what changed, and the bug that hides in itA load on "updated_at greater than the last run" that silently misses rows, and logical replication as the fix.Coming
- Writing back without making duplicatesCreating tasks in Bramfield's work queue so that a timeout and a retry never create two of anything.Coming
- No public internet: reaching systems privatelyReaching Bramfield's systems through a private subnet, VPC endpoints and Session Manager, instead of open ports and SSH keys.Coming
- The corporate proxy and the certificate errorEvery outbound call goes through a proxy that inspects TLS; you reproduce CERTIFICATE_VERIFY_FAILED on a real model call and fix it with the customer's certificate bundle.Coming
- Keeping and rotating the customer's secretsStoring credentials in Secrets Manager, reading them at run time, and rotating one without the service going down.Coming
Part 4
From messy data to safe data
Cleaning the claims data, catching a changed column before it breaks the load, masking personal data before any model sees it, reading claim documents, building the golden set with the claims team, and recording where every number came from.
6 lessons, coming
- Cleaning the claims dataProfiling, cleaning and loading the exports with DuckDB and pandas, and tests that fail loudly instead of loading bad rows.Coming
- When the upstream team changes a columnA renamed column that either breaks the load at 2 a.m. or quietly fills it with nulls, and a data contract that catches it first.Coming
- Finding and masking personal data before the model sees itNames, phone numbers, SSN-format numbers, driver's license numbers and VINs in claim notes, masked with Microsoft Presidio plus recognizers you write, measured on a labeled sample, including the numbers a validity check would wrongly skip.Coming
- Reading claim documents and policy formsTurning PDFs and scanned forms into text and tables you can trust, and checking what was lost on the way.Coming
- Building the golden set with the claims team300 claims with the right answer agreed by Bramfield's claims lead, so every later number means something.Coming
- Where every number came fromRecording lineage so you can answer "which file and which rows did this decision use?" months later.Coming
Part 5
The pilot: an AI that does the job
Choosing the model and where it runs, decisions your code can check, answers from the policy form, safe questions about the data, the claims triage agent and its MCP server, an adjuster approving every draft, a hidden instruction, and an honest demo.
10 lessons, coming
- Choosing the model and where it runsA provider's API with no data kept, a model served inside AWS in us-east-1, or one you host, compared on data location, cost per claim in dollars and speed, with the same prompt.Coming
- A claim decision your code can checkThe model returns a structured decision checked with Pydantic, never free text, and a wrong shape is caught before anyone sees it.Coming
- Answering from the policy formRetrieval over Bramfield's policy forms with pgvector on RDS, with the clause cited in every answer.Coming
- When the right clause is not foundHybrid search and reranking, measured as recall on the golden set before and after.Coming
- Questions about the claims data, answered safelyText-to-SQL through a read-only database role, with every query checked against an allowlist by sqlglot.Coming
- The claims triage agentA LangGraph agent that reads a claim, checks cover, looks for fraud signals and drafts a decision note, using only read-only tools.Coming
- Bramfield's systems as an MCP serverThe policy service, the claims database and the work queue wrapped as one MCP server, so any agent uses the same checked tools.Coming
- A person approves every decisionPausing the agent with interrupt so an adjuster approves, edits or refuses each draft, and the code that re-checks the claim before acting.Coming
- An instruction hidden in a claim documentA claim note that tells the agent to approve the payout, a simple agent that obeys it, and the layered defense that holds.Coming
- The pilot demo that tells the truthShowing the head of claims real cases, failures included, and why a polished demo on easy cases loses trust later.Coming
Part 6
Proving it works
Agreeing with the customer what "it works" means, reading the failures, scoring the agent by code on every change, checking an LLM judge against the claims lead, running beside the adjusters in shadow mode, and presenting the results honestly.
6 lessons, coming
- What "it works" means for BramfieldAgreeing the measures with the customer before testing: correct decisions, wrong payouts in dollars, missed fraud, days to settle and cost per claim.Coming
- Reading the failuresError analysis on 100 real traces from your agent, grouped into the few failure types that matter.Coming
- The evaluation harnessRunning the agent on the golden set and scoring it by code, every time a prompt, a model or a tool changes.Coming
- A judge for the decision notes, checked against the claims leadAn LLM judge for the written notes, measured against the claims lead's own grades before it is trusted.Coming
- Shadow mode: the AI works beside the adjustersTwo simulated weeks where the agent decides every claim silently, compared with what the adjusters actually did.Coming
- The pilot readoutWriting and presenting the results to Bramfield's executives with confidence ranges, the cost and the savings in dollars, and the failures stated plainly.Coming
Part 7
Production inside the customer's walls
The production stack in the customer's own AWS account, private services, traces that never leave, a cost line for every claim, provider failures rehearsed, a rollout one state at a time, debugging without a shell, the security review and an audit trail an examiner can read.
9 lessons, coming
- Building in the customer's AWS accountThe production stack in Terraform, in us-east-1, in an account Bramfield owns, with roles their team can audit.Coming
- Running the service in private subnetsContainers on ECS Fargate behind an internal load balancer, with no public address anywhere.Coming
- Traces and logs that never leave the customerOpenTelemetry to a collector inside Bramfield's network, with personal data dropped before anything is stored.Coming
- Cost per claim and the model gatewayOne gateway for every model call: limits per team, caching, a fallback model, and a cost line in dollars for every claim.Coming
- When the model provider failsTimeouts, rate limits and a full outage, rehearsed on purpose, with the claims queue still working by hand.Coming
- Rolling out to one state firstA feature flag per state, a small first group of adjusters, and a rollback you have actually practised.Coming
- Debugging in their environment without a shellThe 403, the name that will not resolve inside the VPC, and the call that is slow only there, solved with Session Manager, logs and traces.Coming
- Passing the security reviewAnswering Bramfield's vendor security questionnaire with evidence from your own stack: encryption, access, retention, where the data lives, and what a SOC 2 report would and would not cover.Coming
- An audit trail an examiner can readEvery decision, approval and model call recorded in a log that nobody can edit, and the query that answers a state insurance examiner's question.Coming
Part 8
Handover, adoption and the next engagement
Getting the claims team to use it, handing over a system they can run, the field report back to product, turning one deployment into a pattern, the expansion proposal, a capstone engagement you run yourself, and your portfolio and the interview.
7 lessons, coming
- Getting the claims team to use itTraining adjusters, measuring real use from the logs, and fixing the reasons people go back to the old way.Coming
- The runbook and the handoverHanding Bramfield's team a system they can run: the runbook, on-call notes and one failure rehearsed with them.Coming
- The field report back to productWhat you built that only Bramfield needs, what every insurer will need, and how to write it so the product team acts on it.Coming
- Turning one deployment into a patternPackaging the connectors, evals and agent so a second team starts in days, and measuring how long the second start really takes.Coming
- The expansion proposalThe business case for homeowners claims, in dollars, built from the costs and results you measured in the auto pilot.Coming
- Capstone: a new department, end to endAfter a hailstorm, Bramfield's homeowners claims team asks for help; you run discovery, a pilot, evals and a rollout yourself, then tear it all down.Coming
- Your portfolio and the FDE interviewTurning the Bramfield work into a portfolio, and practising the decomposition, customer, debugging and design rounds.Coming