Forward Deployed Engineer — Study Notes
Companion notes mapping an enterprise AI Forward Deployed Engineer curriculum (14 modules + 2 capstones) against this question bank.
Every code example in these notes was executed and verified before publication — measured output is shown rather than described.
Start here
Study Guide & Bank Mapping Which bank sections cover which modules, where the gaps are, and a suggested study order. 51% of this bank already covers the syllabus — the guide shows which half you can skip.
Gap Analysis What the syllabus omits, what the bank omits, and the eight things neither covers — client-environment constraints, the bespoke↔product tension, commercial literacy, and capability handover.
Detailed module notes
| Part | Modules | Covers |
|---|---|---|
| 1 | 1–4 | Python & async, FastAPI & GraphQL, AWS networking & IAM, Docker & CI/CD |
| 2 | 5–8 | LLM fundamentals, vector search & RAG, Neo4j & Cypher, multimodal & ColPali |
| 3 | 9–11 | LangGraph, agent orchestration & MCP, SOAP/XML & legacy SQL |
| 4 | 12–14 | OAuth & data-level RBAC, guardrails & PII, gateways & observability |
| 5 | Capstones | Discovery workshops, SOWs, ROI models, UAT runbooks, MCP trust boundaries |
A few things worth knowing before you start
Post-filtering entitlements returns nothing. In a measured simulation with a user entitled to 3% of a corpus, post-filtering a top-10 search returned zero results — while having already traversed 970 documents the user could not see. Both a quality bug and a security failure. (Part 4)
Fixed backoff synchronises a retry storm. Twenty clients all waited exactly 4.00s (stdev 0.00). Full jitter spread them across 0.01–3.78s. (Part 4)
A 7B model needs 112GB before activations. Weights 14GB + gradients 14GB + FP32 master 28GB + two Adam moments at 28GB each — roughly 16 bytes per parameter. This is why LoRA exists. (Part 1, and bank Q1374)
Two of three PII regex matches were false positives until a Luhn check was applied — which is why “evaluating false-positive redaction rates” belongs in the curriculum. (Part 4)
Automating 90 minutes of a 3-day process changes nothing. The AuditMesh workflow map finds the bottleneck is a human approval, not the analysis. Saying so is the consulting insight. (Part 5)
These notes complement the question bank: the bank covers concepts and architecture at interview depth, while these cover tool fluency and the consulting lifecycle the bank does not.