THE FDE VIDEO LIBRARY

FDE interview videos
& course lessons.

Prepare for forward deployed engineer interviews with animated concept explainers, worked answers and full course lessons. Choose a topic for a quick explanation, or follow a course lesson by lesson. Each video links to its written page.

Videos on this site follow the access requirements of their written pages. Premium lessons require Premium access.

210 video lessons · 24h 13m in total · English narration

Subscribe on YouTube

Foundations of LLMs & GenAI

Explore this topic →

Tokenization & Tokens

Watch the explainer

The Context Window

Watch the explainer

Embeddings & Vector Representations

Watch the explainer

The Transformer, Intuitively

Watch the explainer

Temperature, Top-p and Sampling

Watch the explainer

Prompt Engineering

Watch the explainer

Why LLMs Hallucinate

Watch the explainer

Fine-tuning vs RAG vs Prompting

Watch the explainer

Autoregressive Decoding

Watch the explainer

Dot product vs cosine similarity: when magnitude quietly breaks your retrieval

Watch the explainer

Base model vs instruct model: the difference, and why a customer should care

Watch the explainer

Retrieval & Agents

Explore this topic →

Retrieval-Augmented Generation (RAG)

Watch the explainer

Vector Databases

Watch the explainer

AI Agents and Tool Use

Watch the explainer

TF-IDF and BM25

Watch the explainer

Context Window Management for FDE Agents

Watch the explainer

Walk me through a RAG pipeline: the spine, and where it actually breaks

Watch the explainer

Chunking documents for RAG: why chunk boundaries are a recall problem

Watch the explainer

When is an AI agent the wrong answer? The cost ladder and the three conditions

Watch the explainer

How to evaluate a RAG system before launch: golden set, split metrics, launch gate

Watch the explainer

Hybrid search design: BM25, vectors, RRF and a reranker, and what each stage rescues

Watch the explainer

A RAG pilot answers 60% correctly: the diagnostic tree that wins the question

Watch the explainer

Designing a customer-support agent's tools: reads, two-phase writes, and evals first

Watch the explainer

Evaluation & ML Foundations

Explore this topic →

Precision, Recall and F1

Watch the explainer

Gradient Descent & Learning Rate

Watch the explainer

Bias-Variance Tradeoff

Watch the explainer

Overfitting and Regularization

Watch the explainer

Golden Datasets and Eval Sets

Watch the explainer

Loss Functions

Watch the explainer

Activation Functions

Watch the explainer

Neural Network Basics: Perceptron to MLP

Watch the explainer

ROC AUC vs PR AUC: why a 0.95 AUC model can flood the review queue with junk

Watch the explainer

Class imbalance: 99.5% accuracy that catches nothing, and the full fix

Watch the explainer

Your model scored 95% in the pilot and 70% in production: data leakage, diagnosed

Watch the explainer

K-fold cross-validation, and the two data shapes where it lies to you

Watch the explainer

"We want the model to be accurate": turning that into the metric you optimize

Watch the explainer

Tree ensembles vs linear models vs neural nets: choosing for tabular data

Watch the explainer

System Design for AI in Production

Explore this topic →

From Proof-of-Concept to Production

Watch the explainer

Retries, Exponential Backoff and Jitter

Watch the explainer

At-least-once vs exactly-once delivery: why duplicates are certain, and the fix

Watch the explainer

Rate limiting explained: fixed window, sliding window, token bucket, and when each breaks

Watch the explainer

p50, p95, p99 latency explained: why the tail hits far more than 1% of users

Watch the explainer

Dead-letter queues and poison messages: what happens after the message lands

Watch the explainer

Estimate the capacity and cost of an LLM app: tokens, peak QPS, and tokens per minute

Watch the explainer

A customer's app times out calling your API and you can't see their code: debug it

Watch the explainer

Why a queue consumer double-processes under load, and the race-free fix

Watch the explainer

MLOps & Lifecycle

Explore this topic →

How CI/CD for ML models differs from traditional DevOps CI/CD

Watch the explainer

What MLflow is for, and its four components, with the detail that proves you've used it

Watch the explainer

Why use a model registry instead of just storing model artifacts in S3?

Watch the explainer

What problem a feature store solves: the online store vs the offline store

Watch the explainer

Shadow deployment vs A/B testing a model: different questions, run in sequence

Watch the explainer

ML Infrastructure & Serving

Explore this topic →

The CUDA execution model: how grids, blocks and warps map to SMs

Watch the explainer

The GPU memory hierarchy: registers, shared memory, L2, HBM, and the numbers that matter

Watch the explainer

Warp divergence: why mixed branches in a warp cost you, and which ones don't

Watch the explainer

Memory coalescing and shared-memory bank conflicts, and the one-character transpose fix

Watch the explainer

Memory-bound or compute-bound? The roofline model, with an H100's real numbers

Watch the explainer

Data vs tensor vs pipeline parallelism, and how they combine into 3D parallelism

Watch the explainer

Ring all-reduce: how it works, and deriving its communication cost

Watch the explainer

What ZeRO and FSDP actually shard, how much each stage saves, and where gradient checkpointing fits

Watch the explainer

Data & SQL Engineering

Explore this topic →

SQL Window Functions

Watch the explainer

Second-highest salary per department in SQL: DENSE_RANK, and why RANK returns nothing

Watch the explainer

SQL: keep only the latest row per key, rerun-safe (and the MAX bug that invents rows)

Watch the explainer

Why your LEFT JOIN turned into an INNER JOIN: the WHERE clause NULL trap

Watch the explainer

Gaps and islands in SQL: find 3+ consecutive days with the date-minus-row_number trick

Watch the explainer

Sessionize events in SQL: LAG, a 30-minute gap, and a running SUM

Watch the explainer

Funnel analysis in SQL by signup cohort: count users, enforce order, and the MIN trap

Watch the explainer

SQL join fan-out: why your revenue tripled after one join, and the 30-second check

Watch the explainer

AI Security, Privacy & Governance

Explore this topic →

Direct vs indirect prompt injection: why the indirect one breaks enterprise LLM apps

Watch the explainer

Jailbreak vs prompt injection: whose policy is violated, and who has to fix it

Watch the explainer

Why system prompts leak, and what should never be in one

Watch the explainer

The OWASP LLM Top 10 for AI agents: which two risks to prioritize, and why

Watch the explainer

Excessive agency (OWASP LLM06): the controls that actually limit an AI agent

Watch the explainer

MCP tool poisoning: why tool descriptions are an attack surface, and how to defend them

Watch the explainer

Coding & Engineering Craft

Explore this topic →

Parsing Messy, Real-World Data

Watch the explainer

Big-O That Actually Matters

Watch the explainer

Sliding Window and Two Pointers

Watch the explainer

Heaps and Top-K

Watch the explainer

Two Sum in one pass: the hashmap answer, the trace, and the duplicate trap

Watch the explainer

Group anagrams: the canonical-key idea interviewers are really testing

Watch the explainer

Merge overlapping intervals: sort, then sweep, and the two details interviewers wait for

Watch the explainer

Subarray sum equals K: why sliding window fails, and the prefix-sum hashmap

Watch the explainer

Three Sum: sort, two pointers, and the three duplicate skips interviewers grade

Watch the explainer

Design an LRU cache with O(1) get and put: OrderedDict, then from scratch

Watch the explainer

The Customer-Facing Craft

Explore this topic →

Requirements Discovery

Watch the explainer

Scoping Ambiguous Problems

Watch the explainer

Explaining Trade-offs to Non-Engineers

Watch the explainer

"Your most ambiguous project": how to answer the week-one follow-up

Watch the explainer

"A time you said no to a customer": the yes-shaped no, with a worked example

Watch the explainer

"A project that failed. Whose fault was it?" How to answer with honest attribution

Watch the explainer

Fermi estimate: LLM tokens per day for a Fortune-500 support org, worked through

Watch the explainer

"Turning a skeptic into a champion": steel-man first, then convert with behavior

Watch the explainer

ML System Design (Product)

Explore this topic →

Design the ranking model for a personalized feed: the ML system design interview

Watch the explainer

Design an evaluation framework for an ads-ranking system

Watch the explainer

Design a system to detect bots and inauthentic accounts in real time

Watch the explainer

Build a fraud-detection model for payments: threshold on dollars, not probability

Watch the explainer

Foundations of Forward Deployed Engineering

Explore this topic →

FDE Foundations 01 · The last mile: connect a capability to a real customer workflow

Watch the explainer

The last mile: what a forward deployed engineer actually does

Watch the explainer

FDE Foundations 02 · Roles by ownership: compare FDE, solutions, services and product work

Watch the explainer

FDE vs solutions engineer vs professional services: compare the work, not the title

Watch the explainer

FDE Foundations 03 · The engagement loop: learn as you go, with explicit release gates

Watch the explainer

The FDE engagement loop: discover, scope, evaluate, ship, integrate, observe, learn, productize

Watch the explainer

FDE Foundations 04 · The stated problem: test a customer's request against the real work

Watch the explainer

The stated problem is not always the problem: discovery for forward deployed engineers

Watch the explainer

FDE Foundations 05 · The people who can stop you: map decisions, owners and prerequisites

Watch the explainer

The people who can stop your deployment: sponsors, champions, operators and reviewers

Watch the explainer

FDE Foundations 06 · The scoping document: a scope you can test, approve and change

Watch the explainer

The pilot scoping document: six sections that keep the gaps visible

Watch the explainer

FDE Foundations 07 · Orienting in someone else's system: build an evidence map

Watch the explainer

Reading an unfamiliar system: follow one record, then verify what runs

Watch the explainer

FDE Foundations 08 · Boundaries before components: ownership, trust, network and meaning

Watch the explainer

Boundaries before components: ownership, meaning, trust and network

Watch the explainer

FDE Foundations 09 · Three failure modes: test replay, slow dependencies and changed data

Watch the explainer

Three failure modes the demo misses: repeats, slow dependencies, changed data

Watch the explainer

FDE Foundations 10 · The seam: translate source data without losing its meaning

Watch the explainer

The seam: write the adapter contract before the conversion

Watch the explainer

FDE Foundations 11 · Failing at 3 a.m.: retries from outcome evidence and a time budget

Watch the explainer

FDE Foundations 12 · Tests you can leave behind: catch regressions, state their limits

Watch the explainer

FDE Foundations 13 · The cleaning is the project: profile data into a scoped plan

Watch the explainer

FDE Foundations 14 · Safe to run again: replays that never overwrite newer facts

Watch the explainer

FDE Foundations 15 · Integrating the undocumented: investigate with bounded evidence

Watch the explainer

FDE Foundations 16 · One thing, three systems: resolve identities without hiding the unmatched

Watch the explainer

FDE Foundations 17 · Tables an analyst can use: declare the grain, prove totals survive joins

Watch the explainer

FDE Foundations 18 · When the business changes its mind: valid history and what you knew

Watch the explainer

FDE Foundations 19 · One question: give a proof of concept a decision and an evidence standard

Watch the explainer

FDE Foundations 20 · Fake it honestly: prototype shortcuts that don't change the claim

Watch the explainer

FDE Foundations 21 · Where they can see it: a bounded trial in users' hands

Watch the explainer

FDE Foundations 22 · Retrieval is what fails: trace the evidence before changing the prompt

Watch the explainer

FDE Foundations 23 · Write the eval first: labeled cases, abstentions and honest denominators

Watch the explainer

FDE Foundations 24 · Design for being wrong: abstention, review queues and honest limits

Watch the explainer

FDE Foundations 25 · Config and secrets: validate before serving, test credential rotation

Watch the explainer

FDE Foundations 26 · The three signals: connect logs, metrics and traces to outcomes

Watch the explainer

FDE Foundations 27 · A service level you can defend: denominators, error budgets, policy

Watch the explainer

FDE Foundations 28 · Write for one paragraph: lead with the decision, keep the evidence

Watch the explainer

FDE Foundations 29 · A demo is a narrative: visible evidence, honest limits, one decision

Watch the explainer

FDE Foundations 30 · Handing over: evidence, rehearsal and accepted ownership

Watch the explainer

The FDE Engagement

Explore this topic →

FDE Engagement 01 · The engineer in the sales call: commit only to what you can deliver

Watch the explainer

The engineer in the sales call: establish what you can actually deliver

Watch the explainer

FDE Engagement 02 · Find the deal killer: test what can invalidate a pilot before building

Watch the explainer

Find the deal killer: test the constraints that can sink a pilot before you build

Watch the explainer

FDE Engagement 03 · The pilot proposal: write it around a decision you can actually test

Watch the explainer

FDE Engagement 04 · The system nobody owns: choose an integration path you can operate

Watch the explainer

Integrating the system nobody owns: file drops, APIs, replicas and change feeds

Watch the explainer

FDE Engagement 05 · Proving who you are: enforce the caller's permissions at every boundary

Watch the explainer

FDE Engagement 06 · Pipelines that survive: replay, quarantine and reconciliation

Watch the explainer

FDE Engagement 07 · Objects, not tables: give business definitions an executable home

Watch the explainer

FDE Engagement 08 · The hard part is the links: evidence, uncertainty and history in SQL

Watch the explainer

FDE Engagement 09 · Actions are the write path: authorize, retry and recover safely

Watch the explainer

FDE Engagement 10 · Where it runs is a decision: a deployment you can actually operate

Watch the explainer

FDE Engagement 11 · Reproducible or it didn't happen: recreate the release, recover the state

Watch the explainer

FDE Engagement 12 · Seeing inside a box you can't open: diagnostics without your login

Watch the explainer

FDE Engagement 13 · The review is a conversation: a security packet that supports a decision

Watch the explainer

FDE Engagement 14 · The findings you can predict: follow data copies through deletion

Watch the explainer

FDE Engagement 15 · What they will ask about the model: data paths and enforced tools

Watch the explainer

FDE Engagement 16 · The eval harness comes first: an evaluation that can block a release

Watch the explainer

FDE Engagement 17 · The budget decides the architecture: cost per unit of the customer's work

Watch the explainer

FDE Engagement 18 · Nothing alerts when quality drops: read overrides, mix and coverage

Watch the explainer

FDE Engagement 19 · Rule, model or prompt: write the rule first, then compare on cost

Watch the explainer

FDE Engagement 20 · The training set is the project: leakage, availability time, splits

Watch the explainer

FDE Engagement 21 · It decays and nothing errors: investigate drift before you retrain

Watch the explainer

FDE Engagement 22 · Authority without blast radius: bound agent actions and shared limits

Watch the explainer

FDE Engagement 23 · What a half-finished run leaves behind: recover agent runs without guessing

Watch the explainer

FDE Engagement 24 · You promote an agent, you don't deploy it: evidence gates for agent autonomy

Watch the explainer

FDE Engagement 25 · The number they will quote back: baseline, time, capacity and cash

Watch the explainer

FDE Engagement 26 · Attribution without a control group: improvement is not causation

Watch the explainer

FDE Engagement 27 · The review meeting: write the memo around a decision the evidence supports

Watch the explainer

FDE Engagement 28 · The champion who leaves: make the engagement survive a change of sponsor

Watch the explainer

FDE Engagement 29 · Saying no without losing the room: evidence, choices and boundaries

Watch the explainer

FDE Engagement 30 · The engagement outlives you: transfer operation, decisions and change

Watch the explainer

Production AI Agents for FDEs

Explore this topic →

Production AI Agents 01 · Choose the agent boundary: put the model where uncertainty belongs

Watch the explainer

Production AI Agents 02 · Define success before prompts: grade outcomes, keep missing cases

Watch the explainer

Production AI Agents 03 · Retrieve the applicable policy: hard filters before relevance ranking

Watch the explainer

Production AI Agents 04 · Compile a bounded context: evidence the model can safely use

Watch the explainer

Production AI Agents 05 · Design read tools for decisions: narrow tools, server-bound identity

Watch the explainer

Production AI Agents 06 · Bind approval to the action: exact proposals, rechecked at the write

Watch the explainer

Production AI Agents 07 · Persist intent before sending: action identity and atomic receipts

Watch the explainer

Production AI Agents 08 · Recover an unknown outcome: timeouts, receipts and reconciliation

Watch the explainer

Production AI Agents 09 · Read an agent evaluation: paired cases, missing outcomes, release gates

Watch the explainer

Production AI Agents 10 · Calibrate the reviewer: rubrics, disagreements and exact checks

Watch the explainer

Production AI Agents 11 · Budget the whole support case: failed attempts and review capacity

Watch the explainer

Production AI Agents 12 · Release with a working stop: staged capabilities and recovery

Watch the explainer

Hard Deployments

Explore this topic →

Hard Deployments 01 · Does it fit? Weights, KV cache and runtime memory in consistent units

Watch the explainer

Does the model fit? Weights, KV cache and overhead in GiB, worked through

Watch the explainer

Hard Deployments 02 · Throughput is not latency: queueing, prefill and decode, measured apart

Watch the explainer

Hard Deployments 03 · Sizing before they buy: a capacity estimate that survives the purchase meeting

Watch the explainer

Hard Deployments 04 · What your build assumes about the internet: offline dependencies, tested

Watch the explainer

Hard Deployments 05 · Shipping through a one-way door: recovery inside the boundary

Watch the explainer

Hard Deployments 06 · Evidence a regulator accepts: make it verifiable, keep gaps visible

Watch the explainer

Hard Deployments 07 · What isolation you actually guarantee: specify and test every layer

Watch the explainer

Hard Deployments 08 · The tenant who ruins the median: per-tenant evidence, admission and fairness

Watch the explainer

Hard Deployments 09 · Blast radius and cells: failure scope from dependencies and rollout state

Watch the explainer

Hard Deployments 10 · The model server is a layer too: cache, adapter and shared-failure boundaries

Watch the explainer

Hard Deployments 11 · Where the money actually goes: reconcile the bill to attempts and outcomes

Watch the explainer

Hard Deployments 12 · The three levers: cache, send less, or route cheaper without losing quality

Watch the explainer

Hard Deployments 13 · Planning against a price ceiling: fixed costs, margin targets and workload mix

Watch the explainer

Hard Deployments 14 · State that survives the process dying: durable intent and unknown outcomes

Watch the explainer

Hard Deployments 15 · Memory that does not rot: freshness, conditional writes and compaction

Watch the explainer

Hard Deployments 16 · A fleet of agents, one incident: mix-driven alarms and stop mechanisms

Watch the explainer

Hard Deployments 17 · The first 48 hours: contain harm, grade evidence, forecast conditionally

Watch the explainer

Rescuing a failing deployment: the first 48 hours, from containment to assessment

Watch the explainer

Hard Deployments 18 · Proving what happened: forensic queries, denominators and bounded findings

Watch the explainer

Hard Deployments 19 · The fourteen-day plan: dependency paths, residual work and honest forecasts

Watch the explainer

Hard Deployments 20 · Service levels across a portfolio: burn rates, empty traffic and change policy

Watch the explainer

Hard Deployments 21 · Incident command without access: bounded requests, packets and timelines

Watch the explainer

Hard Deployments 22 · Degradation by design: fallbacks, valid zeros and recovery evidence

Watch the explainer

Hard Deployments 23 · Finding the abstraction: share mechanisms, keep policy in adapters

Watch the explainer

Hard Deployments 24 · Configuration over customisation: validate settings, test behavior

Watch the explainer

Hard Deployments 25 · Migrating off the one-off: shadow comparison, exact approvals, real rollback

Watch the explainer

Hard Deployments 26 · Evaluation infrastructure for a fleet: one harness, per-deployment cases

Watch the explainer

Hard Deployments 27 · The field-to-product loop: count, price and propose the change

Watch the explainer

Hard Deployments 28 · So it does not run through you: transfer capability, authority and judgment

Watch the explainer

Turn a video lesson into interview practice.

Pause a worked example and explain the next step yourself. Then try the linked interview questions or continue through the FDE courses with their exercises.

Practice questions →