Free course · 11 lessons · Advanced
Forward Deployed Engineering in Practice
What goes wrong when software meets real customers, and how forward deployed engineers fix it: messy production data, corporate proxies and TLS inspection, rate-limited backfills, prototypes that can't ship, personal data in logs, SSO launch-day failures, AI regressions without evals, retrieval that serves stale documents and pilots drowned in requests, with runnable simulations and two capstones.
1. Integrations That Break
Integrations work in the demo and fail in the field: real data, corporate networks and other people's rate limits.
- Works on the Sample, Fails on Real Data — Customer sample files are clean; production exports aren't. Validate, quarantine and report instead of crashing or silently dropping rows.
- It Works From Our Office — Inside the customer's network every call fails with certificate errors or timeouts. Support proxies, custom CAs and allow-lists from day one.
- The Backfill That Got Us Blocked — An initial historical sync hammered the customer's API, triggered rate limits and got the integration account locked. Throttle, back off, checkpoint and agree backfills in advance.
2. Shipping in Customer Environments
Prototypes become production, logs leak data, and sign-in breaks on launch day. Make the jump to production deliberately.
- The Demo That Couldn't Ship — A notebook prototype won the deal and then couldn't pass the customer's go-live review. Plan the production gap from the start and close it in small steps.
- Customer Data in Our Logs — Debug logging copied bank account numbers and emails into a third-party logging service. Log identifiers, redact at the source, and agree what may leave the customer.
- Locked Out on Launch Day — On go-live morning nobody could sign in: clock skew, a wrong audience and missing group mappings. Test SSO end to end, allow small skew, and keep a break-glass admin.
3. AI Deployments in the Field
AI deployments fail quietly: a prompt tweak regresses accuracy, retrieval serves stale documents, and the pilot drowns in requests.
- The Prompt Change That Broke Extraction — A small prompt edit fixed one customer complaint and silently broke date extraction for thousands of documents. Build a golden set and gate every change on it.
- Confidently Wrong Answers From the Wiki — An internal assistant cited a policy that was replaced two years ago. Fix retrieval: curate sources, add metadata filters, rank fresh documents, and say 'I don't know'.
- Twelve Requests, No Priorities — A promising pilot stalls under a flood of requests from every stakeholder. Anchor on success criteria, triage openly, and turn patterns into product feedback.
4. Capstone Case Studies
Plan an AI deployment at a regulated customer end to end, and diagnose a pilot that stalled.
- Capstone: Deploying an AI Document Assistant at a Bank — A bank wants an assistant that drafts answers to customer complaints from policies and case history, in 12 weeks. Plan the deployment end to end.
- Capstone: Postmortem — The Pilot That Stalled — A 10-week pilot ended without a production decision. Reconstruct what went wrong and how a forward deployed engineer could have changed the outcome.