Awesome-FDE-Roadmap

The Awesome-FDE-Roadmap, a comprehensive guide for Forward Deployed Engineers (FDEs), outlines the hybrid role of software engineer, AI/data architect, and strategic consultant, emphasizing the need to bridge the gap between core products and client realities. The roadmap covers a master curriculum including advanced SQL, data modeling, distributed computing, and cloud networking, with recommended tools such as Python, Go, SQL, dbt, DuckDB, Apache Spark, Terraform, Helm, GCP, Prometheus, Grafana, and Loki. It positions FDEs as critical for ensuring multi-million dollar contracts succeed by architecting solutions, managing stakeholders, and writing glue code in hostile, legacy, or air-gapped environments.

Forward Deployment Engineering is the art of shipping complex software where it matters most. It is a hybrid role—part Software Engineer , part AI/Data Architect , and part Strategic Consultant . FDEs are the "Technical Special Ops" who bridge the gap The Delta between a core product and a client’s messy, real-world reality. The FDE Persona & Mission The Master Curriculum The Applied AI & Technical Playbook The "Soft Stack": Consulting & Strategy The Interview Blackbook & Case Studies Artifact Templates Copy-Paste Comprehensive Reading List The FDE Glossary Creators Contributing The FDE is the bridge between the "Perfect Code" of HQ and the "Messy Reality" of the Client. The Problem: Standard software often fails when it hits real-world data corrupt schemas, air-gapped servers, political resistance . The Solution: The FDE. They don't just "fix bugs"—they architect solutions, manage stakeholders, and write the "glue code" that makes a multi-million dollar contract successful. In traditional Software Engineering SWE , you build for a "user persona". In FDE, you build for a mission . | Feature | Software Engineer SWE | Forward Deployed Engineer FDE | |---|---|---| User | Millions of anonymous users | High-stakes stakeholders CTOs, Generals, CEOs | Environment | Controlled, uniform cloud | Hostile, legacy, air-gapped, or hybrid infra | Goal | Scale and stability | Speed-to-value and problem-solving | Code Ratio | 90% Features | 50% Integration/Glue, 50% Strategy | Languages: Python Data/AI , Go Infra , SQL Everything . Data: dbt Transformation , DuckDB Local processing , Apache Spark . Cloud: Terraform IaC , Helm K8s packaging , GCP . Observability: Prometheus , Grafana , Loki . The FDE mission usually starts with a "Data Audit". If you can't untangle a client’s 20-year-old schema, you can't build on top of it. Advanced SQL & Query Tuning: Beyond JOINs . You must master Window Functions, Recursive CTEs, and Query Optimization. You should be able to look at an EXPLAIN plan and identify why a query is scanning 10TB of data unnecessarily. Data Modeling for Reality: Understand Star Schema vs. One Big Table OBT . Learn to design schemas that balance write-performance with user-readability. The Medallion Architecture: Bronze: Raw landing zone immutable . Silver: Filtered, joined, and cleaned the "Single Source of Truth" . Gold: Business-ready aggregates Powering the UI/AI . Distributed Computing: Understanding Spark/Ray partitioning. You must know how to debug "Data Skew" and "OOM Out of Memory " errors when processing client datasets that exceed local memory. Data Quality & Observability: Implementing "Circuit Breakers" for data. If a client's upstream data breaks, your pipeline should alert you before the CEO sees a broken dashboard. An interactive book to master SQL through real-world data. Select Star SQL https://selectstarsql.com/ :The "Bible" for understanding how data systems actually work under the hood. Designing Data-Intensive Applications DDIA https://www.oreilly.com/library/view/designing-data-intensive-applications/9781491903063/ :The standard for turning "Data Engineering" into "Analytics Engineering". dbt Fundamentals https://courses.getdbt.com/courses/fundamentals :A critical tool for fast, local analysis of client CSV/Parquet files without setting up a full cluster. DuckDB for FDEs https://duckdb.org/docs/ : FDEs are often "dropped" into complex cloud environments. You must move beyond writing code to architecting the "Landing Zone" where that code lives. Using GCP as our standard, you must master how to securely deploy and scale within a client's project. Cloud Networking & Security VPC Mastery : The Global VPC: Understanding GCP’s unique global networking. Mastering Shared VPCs to allow multiple teams to use a single network. Secure Connectivity: Implementing Cloud Interconnect or Cloud VPN to bridge a client's on-prem data center to GCP. Zero-Trust: Using Identity-Aware Proxy IAP to grant access to internal apps without a VPN. Kubernetes as a Standard GKE : GKE Autopilot vs. Standard: Knowing when to trade control for operational ease. Workload Identity: The gold standard for security—allowing GKE service accounts to act as IAM service accounts without managing JSON keys. Private Clusters: Deploying GKE without public IP addresses to satisfy strict enterprise security requirements. Data Architecture on GCP: BigQuery Integration: Designing schemas for BigQuery Clustering vs. Partitioning to handle petabyte-scale analysis for clients. Serverless Pipelines: Using Cloud Functions or Cloud Run for lightweight, event-driven data processing. Pub/Sub: Architecting real-time streaming "glue" between client systems and your platform. Data Exfiltration Prevention VPC Service Controls : - FDEs in high-security sectors Finance/Gov must master VPC SC . This defines a security perimeter around Google-managed services to prevent data from being moved to unauthorized projects. - FDEs in high-security sectors Finance/Gov must master Infrastructure as Code Terraform : - Automating the entire "FDE Environment". If you can’t spin up a GKE cluster, a BigQuery dataset, and an IAM policy in 5 minutes via Terraform, you aren't ready to deploy "forward". The definitive guide to building secure, resilient, and cost-effective systems on GCP. Google Cloud Architecture Framework https://cloud.google.com/architecture/framework :Essential reading for understanding how traffic flows inside a cluster. GKE Networking Deep Dive https://cloud.google.com/kubernetes-engine/docs/concepts/network-overview :The documentation you will live in daily. Terraform Provider for Google Cloud https://registry.terraform.io/providers/hashicorp/google/latest/docs :Hands-on labs for BigQuery and Dataflow. Google Cloud "Skills Boost" Data Engineer Path https://www.cloudskillsboost.google/paths/16 :A critical skill for FDEs working with sensitive enterprise data. VPC Service Controls VPC SC Explained https://cloud.google.com/vpc-service-controls/docs/overview :Specifically the "Monitoring" and "Incident Response" chapters to keep client deployments alive. Google SRE Workbook https://sre.google/workbook/table-of-contents/ : An FDE is a "Technical Diplomat". You must solve the people problems to allow the technical solutions to work. Structured Problem Solving MECE : Mutually Exclusive, Collectively Exhaustive. Break a massive "AI Strategy" into small, non-overlapping technical tasks. The Pyramid Principle: Start with the answer first. Executives want the "So What?" before the technical "How". 80/20 Value Scoping: Identify the 20% of features that will solve 80% of the client's pain. Avoid "Gold-Plating" building complex features no one asked for . The Art of the Discovery Call: Learning to ask "Why?" five times to find the root cause of a business problem. Requirements Translation: Turning a CEO's "I want more efficiency" into a Jira Ticket for "Optimizing Route Pathfinding Algorithm". How to move from being a "vendor" to a "partner". The Trusted Advisor https://trustedadvisor.com/books/the-trusted-advisor :The gold standard for executive communication. The Pyramid Principle Summary https://medium.com/lessons-from-mckinsey/the-pyramid-principle-f0885dd3c5c7 :Sounds cliché, but vital for dealing with resistant IT staff at client sites. How to Win Friends and Influence People https://www.amazon.com/How-Win-Friends-Influence-People/dp/0671027034 :Frameworks for thinking about business problems like an engineer. The McKinsey Way https://www.amazon.com/McKinsey-Way-Ethan-M-Rasiel/dp/0070534489 : As an FDE, you are the bridge between "State-of-the-Art" research and "Production-Grade" deployment. This section maps the architectures you must master and the Google Cloud tools you will use to ship enterprise-grade value. Read these to understand how elite teams solve the "messy reality" of enterprise deployment. – The mandatory "Origin Story" of the FDE role. Palantir: Dev vs. Delta https://blog.palantir.com/dev-versus-delta-demystifying-engineering-roles-at-palantir-ad44c2a6e87 – Real-world case studies on deploying GPT-4 into complex workflows e.g., Morgan Stanley, Harvey . OpenAI: Customer Stories https://openai.com/customer-stories – Focus on GKE, BigQuery, and Vertex AI enterprise patterns. Google Cloud: Architecture Blog https://docs.cloud.google.com/architecture – Deep dive into the "Evals" mindset required for production AI. Anthropic: Evaluating AI Agents https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents The Agent Development Kit ADK is Google’s open-source, code-first framework for building sophisticated multi-agent systems. It treats agent development like software engineering , providing modularity, hierarchy, and deterministic control. Multi-Agent by Design: Compose specialized agents in a hierarchy e.g., a "Manager" delegating to "Researcher" and "Coder" agents . The Agent2Agent A2A Protocol: An open standard that allows agents to discover and communicate with each other via consistent HTTP-based interfaces. Model Agnostic: While optimized for Gemini, ADK uses LiteLLM to support GPT-4o, Claude, and Mistral. Deployment: Native integration withfor managed, auto-scaling production runtimes. Vertex AI Agent Engine https://cloud.google.com/vertex-ai/docs/generative-ai/agent-engine/overview FDEs don't just "vibes-test" their agents; they use a two-loop evaluation framework to prove reliability to the client. Focuses on fast, manual, and interactive debugging during development. A CLI and Web UI tool to test execution paths against "Golden Datasets". adk eval : Metrics: tool trajectory avg score Did it use the right tools? , response match score ROUGE similarity , and rubric based final response quality . Scalable, automated evaluation for high-volume production data and CI/CD integration. FDEs use this to prove that a model update or a prompt change is a measurable improvement across thousands of test cases. The unified platform for both Vertex AI Gen AI Evaluation Service https://cloud.google.com/vertex-ai/docs/generative-ai/models/evaluate-models : Rapid Evaluation synchronous, for dev/test and Pipeline Evaluation asynchronous, for massive datasets . Pairwise Evaluation The evolution of AutoSxS : A "Model-as-a-Judge" approach. It uses a superior model e.g., Gemini 3 Pro as an autorater to compare two model responses Model A vs. Model B based on a specific rubric, providing win rates and detailed explanations for every "judgment." Pointwise Evaluation The RAG Triad : Assessing single model responses against specific quality dimensions using the Rapid Eval API : Groundedness: Does the response strictly follow the retrieved context? Crucial for eliminating hallucinations . Fulfillment: Did the agent actually follow the instructions in the system prompt? Summarization & Coherence: Evaluating the linguistic quality and density of the output. Essential for "Day 2" operations. FDEs set up monitoring to detect Vertex AI Model Monitoring https://cloud.google.com/vertex-ai/docs/model-monitoring : Prediction Drift and Feature Attribution changes in production, ensuring the agentic system doesn't degrade over time as client data evolves. Ingestion: Usingto extract data from complex enterprise PDFs/tables. LlamaParse https://developers.llamaindex.ai/python/framework/llama cloud/llama parse/ Grounding: Usingas a managed RAG engine for semantic retrieval over client data. Vertex AI Search https://docs.cloud.google.com/generative-ai-app-builder/docs Vector Storage: High-scale indexing with. Vertex AI Vector Search https://docs.cloud.google.com/vertex-ai/docs/vector-search/overview Hybrid Search: Combining semantic vectors with keyword-based BM25 search to satisfy specific industry nomenclature. Master clustering and partitioning for TB-scale client datasets. BigQuery Performance Tuning https://docs.cloud.google.com/bigquery/docs/best-practices-performance-overview :Mandatory for FDEs in Finance/Gov to satisfy data privacy requirements. VPC Service Controls VPC SC https://cloud.google.com/vpc-service-controls/docs/overview :Automating the spin-up of GKE, BigQuery, and Vertex AI environments. Infrastructure as Code Terraform https://registry.terraform.io/providers/hashicorp/google/latest/docs : Tracking agent latency and debugging failed tool calls in the field. Cloud Trace & Logging https://cloud.google.com/stackdriver :Integrated with ADK to visualize exactly where an agent's "chain of thought" broke. LangSmith Tracing https://www.langchain.com/langsmith :– The ultimate resource for architecting systems that don't crash under client load. The System Design Primer https://github.com/donnemartin/system-design-primer – Start here to build your first multi-agent team. Google ADK Quickstart https://github.com/google/adk-python – Production-ready templates with built-in CI/CD and evaluation. Agent Starter Pack https://github.com/GoogleCloudPlatform/agent-starter-pack – Best end-to-end RAG education. Pinecone: RAG Learning Center https://www.pinecone.io/learn/series/rag/ An FDE is a "Technical Diplomat". If you cannot navigate the boardroom, your code will never reach production. In the field, you will face ambiguity . A CEO might say, "I want our AI to make us more efficient". As an FDE, your job is to translate that vagueness into a specific Technical PRD Product Requirements Document . When a client asks for a feature, the FDE uses the "Three Whys" to find the root business pain: "What is the System of Record?" Where is the ground truth data? If it's an Excel sheet on someone's desktop, the project is already at risk. "What is the Cost of Inaction?" If we don't build this, what happens? This defines the project's priority. "What does 'Day 2' look like?" Who maintains this once the FDE leaves? If there is no internal owner, the project will die. Before writing a single line of code, you must clear these hurdles: - The "Champion": Who is the internal person fighting for this project? - The "Blocker": Which department usually IT or Legal is most likely to stop us? - The Success Metric: Is success "Lower Latency," "Higher Accuracy," or "Headcount Reduction"? - Classification: Is the data PII, PHI, or Secret? - Ingestion: Is the data "Streaming" Pub/Sub or "Batch" BigQuery transfers ? - Compliance: Do we need VPC Service Controls or Data Loss Prevention DLP masking? - Access: Do we have Project Editor or Owner roles in the Google Cloud Project? - Connectivity: Is this a Private GKE cluster? Do we need a Cloud VPN or Interconnect? - Quotas: Does the client have enough GPU quota A100/H100 for the models we plan to deploy? FDEs use these McKinsey-style frameworks to structure technical chaos. - The Pyramid Principle https://medium.com/lessons-from-mckinsey/the-pyramid-principle-f0885dd3c5c7 : Bottom-Line Up Front BLUF . When talking to a CTO, give the conclusion first, then the supporting technical data. - Mutually Exclusive, Collectively Exhaustive . Ensure your project plan covers all bases without overlapping work. The MECE Principle https://en.wikipedia.org/wiki/MECE principle : - The Trusted Advisor: The formula:$Trust = \frac{Credibility + Reliability + Intimacy}{Self-Orientation}$ . As an FDE, you must lower your "Self-Orientation" focusing on the client's win, not your product's features . - The Delta Concept: Focus entirely on the "Delta"—the specific bridge between what the product does out-of-the-box and what the client needs it to do to be successful. The documents an FDE creates to ensure project survival. The SOW Statement of Work : This is your shield against "Scope Creep". It defines exactly where your job ends. The MVA Minimum Viable Architecture : Don't over-engineer. Design the simplest version on GCP e.g., Cloud Run + BigQuery that proves the value in < 30 days. The Technical Demo: An FDE demo is not a feature tour. It is a Value Narrative . Show how the data moves from their messy legacy system into a clean AI-driven insight. If you see these during discovery, escalate immediately: "Data will be ready in 2 weeks". It never is. "We don't need a project manager on our side". The project will lose direction. "Can we just run this on-prem for now?" This usually indicates a deep-seated distrust of Cloud/GCP that will block the project later. Essential for learning "Consultant-speak". The McKinsey Way https://www.amazon.com/McKinsey-Way-Ethan-M-Rasiel/dp/0070534489 :Learning to identify the "crux" of a client's problem. Good Strategy / Bad Strategy https://www.amazon.com/Good-Strategy-Bad-Strategy-Difference/dp/0307886239 :A critical tool for FDEs to document Architecture Decision Records ADRs https://github.com/joelparkerhenderson/architecture-decision-record : why a certain design choice was made at a client site. FDE interviews at companies like Palantir, Google, or Scale AI don't just test your coding; they test your "Delta" —your ability to bridge the gap between a product and a mission. When given a case study, do not start coding. Use this four-step diagnostic approach: Clarify: Ask about data volume, security PII/PHI , and the "Definition of Done". Architect: Design the data flow from source system to end-user UI using GCP primitives. Solve The Delta : Identify what the product doesn't do out of the box and how you will build the "glue". Evaluate: How do we prove the AI isn't hallucinating? How do we monitor performance? Scenario: "A massive hospital chain wants to use our software to predict patient readmission. They have 20 years of data in a legacy SQL Server on-prem. They have zero cloud presence and extreme HIPAA privacy concerns. Walk us through your first 30 days". Days 1–7 Discovery & Trust : Technical: Run a data profiling audit on the SQL Server. Identify key features age, diagnosis, last visit . Strategy: Meet with the Chief Medical Officer to define "Readmission" is it 30 days or 90? . Build rapport with the IT team that feels "threatened" by the cloud move. Days 8–15 Secure Landing Zone : Architecture: Propose a GCP Landing Zone . Use Cloud Storage for ingestion and BigQuery for the data warehouse. Security: Implement VPC Service Controls and Sensitive Data Protection DLP to mask PII before it hits the analytics layer. This satisfies the HIPAA requirement. Days 16–25 The Agentic Pipeline : Engineering: Build a pipeline using Vertex AI Search grounded in the patient’s history. The Delta: Write a custom Python service on Cloud Run that pulls real-time patient "vitals" from the SQL Server to update the prediction. Days 26–30 Value Validation : Evaluation: Use AutoSxS to compare the model's predictions against historical outcomes. UAT: Put a simple dashboard in front of 5 doctors. If they don't change their behavior based on the data, the project has failed. Question: "A client has 5PB of data on-prem and needs it in BigQuery in 48 hours for an emergency exercise. How do you do it?" FDE Answer: "Internet bandwidth is the bottleneck. I would request a Google Cloud Transfer Appliance high-capacity storage server to be shipped to the site. While the box is in transit, I’d build the BigQuery schema and partitioning strategy to ensure the data is immediately queryable upon upload". Question: "The client’s Lead Engineer hates our product and refuses to give you VPC access. How do you handle it?" FDE Answer: "This is a trust problem, not a technical one. I’d set up a 1-on-1 to understand their concerns. Often, they fear the product will replace their job. I’d show them how our platform automates the 'grunt work' ETL/Ops , allowing them to focus on high-level architecture. I’d offer to co-author the initial deployment scripts to give them ownership". Question: "A bank wants real-time fraud detection <100ms using an LLM. How do you architect this?" FDE Answer: "An LLM is too slow for the primary path. I’d architect a two-tier system: Use a fast, deterministic model XGBoost/Vertex AI for the 100ms decision. Then, pass the 'flagged' transactions to a Gemini-powered agent via Vertex AI Reasoning Engine for an asynchronous, deep-dive explanation that the fraud analyst can read 5 seconds later". A masterclass in integrating thousands of disparate data sources beds, staff, PPE into a single "Operating System" in weeks. Palantir vs. COVID-19 UK NHS https://www.palantir.com/uk/healthcare/ :How FDEs turned 100,000+ PDFs of financial research into an internal "Assistant" that maintains the bank's strict compliance standards. OpenAI & Morgan Stanley https://openai.com/customer-stories/morgan-stanley :Deploying computer vision models to the "Tactical Edge"—processing satellite and drone data where internet is intermittent. Scale AI & the US Army https://scale.com/blog/scale-ai-dod-expand-army-rd-partnership :Using GCP to modernize manufacturing and deploy AI across the supply chain. Google Cloud & Ford https://corporate.ford.com//articles/products/ford-and-google-to-accelerate-auto-innovation/ : Incorporate a "Case Study Rubric" —a checklist of what a "Senior" vs. "Junior" answer looks like. Junior: Focuses only on the code/Python script. Senior: Mentions security, cost-optimization GCP FinOps , and stakeholder buy-in. In the field, your documentation is your contract. Use these templates to define reality, manage expectations, and prove value. Used in Week 1 to document the "messy reality" before you start building. 🕵️ Site Survey: Client Name - Project Name Date: YYYY-MM-DD | Lead FDE: Your Name 1. The Data Landscape The Ground Truth - Source Systems: e.g., On-prem SQL Server 2016, SAP HANA, Unstructured Sharepoint - Data Gravity: e.g., 5TB total, growing at 50GB/day. Must stay within EU regions. - Known Quality Issues: e.g., Missing timestamps in 30% of records, no primary keys in CRM. 2. Technical & Security Constraints - Identity: e.g., Client uses Okta for OIDC. Need to integrate with GCP IAM. - Connectivity: e.g., No public internet. Must use Cloud Interconnect + Private Google Access. - Exfiltration Risks: VPC Service Controls VPC SC are active. Need to request perimeter bridges. 3. The "Delta" The Gap - Product Gap: Core product doesn't support the client's proprietary .xyz file format. - Proposed Glue: Build a custom GCF Cloud Function parser to convert .xyz to Parquet. 4. The Quick Win Week 2 Objective - e.g., Stand up a Vertex AI Search instance on the 'Policy' dataset to prove 90% retrieval accuracy. The "Contract" that defines what success looks like. 📐 Technical Scoping Document: Feature Name 1. Objective & User Persona Enable User Group, e.g., Risk Analysts to perform Action, e.g., Fraud Investigation by leveraging Technology, e.g., Multi-Agent ADK Swarm . 2. Definition of Success The Evals Success is not "it works"; success is measurable: - Retrieval: 90% Hit Rate on Top-3 documents. - Latency: End-to-end agent reasoning < 5 seconds. - Groundedness: 0% Hallucination rate on Golden Dataset manually verified by Client . 3. Phased Deployment Strategy - Phase 1 MVP : Manual trigger agent on Cloud Run using BigQuery static export. - Phase 2 Scale : Automated trigger via Pub/Sub on real-time data stream. 4. Out of Scope - Integration with the legacy AS400 mainframe deferred to Q3 . A Mermaid/Excalidraw diagram showing a modern, high-scale FDE deployment. graph LR subgraph "Client Environment On-Prem " DB Legacy SQL -- Cloud Interconnect -- B end subgraph "GCP Landing Zone" B Cloud Storage / GCS -- C Vertex AI Search C -- D{Google ADK Agent Engine} subgraph "Multi-Agent Swarm" D -- E Planner Agent D -- F SQL Coder Agent D -- G Reviewer Agent end F -- Query -- H BigQuery E -- Grounding -- C end subgraph "Security Layer" I VPC Service Controls J Identity-Aware Proxy end User User -- J -- D The weekly document that justifies the contract renewal. 🛰️ Weekly Executive Summary: Project Name Reporting Period: Date Range | Status: 🟢 GREEN 🚀 Value Delivered This Week - Metric Move: Reduced manual data lookup time for Analysts by 40% via the new Search Agent. - Milestone: Successfully cleared the Security Review for the GKE Private Cluster. - Ingestion: 1.2B rows of historical logs moved into BigQuery; partitioning optimized for cost. ⚠️ Risks & Strategic Blockers - Risk: Client IT team has delayed the Firewall port opening for the VPN. - Impact: Potential 3-day slide on the "Real-time" dashboard milestone. - Action Required: Need Executive Sponsor Name to approve the exception ticket 12345. 🗓️ The "Day 30" Horizon - Finalize AutoSxS evaluation for the production agent. - Transition 1st-line support to the internal Client Ops team. Being a "Forward" engineer means staying six months ahead of the industry. This list is curated to move you from a "coder" to a "system architect and strategist". - 📗 The "Bible". If you only read one book on this list, make it this one. It explains the Designing Data-Intensive Applications https://www.oreilly.com/library/view/designing-data-intensive-applications/9781491903063/ Martin Kleppmann : why behind every database and distributed system you will use on GCP. - 📘 FDEs fail more often due to broken trust than broken code. This book teaches you how to move from a "vendor" to a "strategic partner". The Trusted Advisor https://trustedadvisor.com/books/the-trusted-advisor David Maister : - 📙 The McKinsey standard for communication. Learn to lead with the conclusion and support it with data—essential for talking to client executives. The Pyramid Principle https://www.amazon.com/Pyramid-Principle-Logic-Writing-Thinking/dp/0273710516 Barbara Minto : - 📕 Essential for Phase 2. It teaches you how to "glue" legacy systems together using messaging, gateways, and translators. Enterprise Integration Patterns https://www.enterpriseintegrationpatterns.com/ Gregor Hohpe : - 📓 FDE is often a "Staff-plus" role in terms of scope. This book helps you navigate the high-level technical leadership required at client sites. Staff Engineer: Leadership beyond the management track https://staffeng.com/book Will Larson : FDEs should understand the "ancestry" of the GCP tools they use. The ancestor of GCS Google Cloud Storage . The Google File System 2003 https://research.google.com/archive/gfs-sosp2003.pdf :The foundation of NoSQL on GCP. Bigtable: A Distributed Storage System for Structured Data 2006 https://research.google.com/archive/bigtable-osdi06.pdf :The paper that started the Transformer/LLM revolution. Attention Is All You Need 2017 https://arxiv.org/abs/1706.03762 :The logic behind how Agentic systems like Google ADK actually work. ReAct: Synergizing Reasoning and Acting in Language Models 2023 https://arxiv.org/abs/2210.03629 : - 🎙 The best podcast for the "AI Engineer" era. Deep dives into RAG, Agents, and Evals. Latent Space https://www.latent.space/ : - 🎙 Interviews with the people actually building the frontier models you will be deploying. The Cognitive Revolution https://www.cognitiverevolution.ai/ : - 🎙 Search their archives for "GCP," "Palantir," or "Distributed Systems". Software Engineering Daily https://softwareengineeringdaily.com/ : - 🎙 Vital for staying updated on the "Modern Data Stack". The Data Engineering Podcast https://www.dataengineeringpodcast.com/ : - 📧 A weekly summary of AI progress and—crucially—AI policy/safety. Import AI Jack Clark https://jack-clark.net/ : - 📧 A non-official but highly curated list of every update in the Google Cloud ecosystem. GCP Weekly https://www.gcpweekly.com/ : - 📧 Deep technical analysis of LLM training and alignment. Interconnects Nathan Lambert https://www.interconnects.ai/ : - 📧 Insights into how big tech companies actually operate and ship software. The Pragmatic Engineer https://blog.pragmaticengineer.com/ : A fantastic archive of CS paper summaries. The Morning Paper Archive https://blog.acolyer.org/ :Incredible deep dives on building production-grade ML and Recommendation systems. Eugene Yan’s Blog https://eugeneyan.com/ :List of useful resources about Data, AI & Cloud. Keeping-Learning https://github.com/pierpaolo28/Keeping-Learning : The Delta: Pioneered by Palantir, this refers to the "gap" between what a product does out-of-the-box and the specific engineering required to make it solve a client's mission. Productized Consulting: The philosophy of solving a client’s unique problem through code that can eventually be abstracted back into the core product features. Embedded Engineering: Unlike a consultant who "advises," an FDE is embedded—they have client credentials, sit in client Slack channels, and ship code directly into the client’s production environment. Last-Mile Integration: The complex work of stitching a modern SaaS/AI platform into legacy, often undocumented, "messy" enterprise systems. Air-Gap / Tactical Edge: Environments with zero or intermittent internet connectivity Common in Defense/Energy . Requires local container registries and offline model weights. VPC Service Controls VPC SC : A GCP security perimeter that prevents data exfiltration by restricting access to Google-managed services like BigQuery or Vertex AI only from authorized networks. Hardening: The process of moving a prototype from "it works on my machine" to "it meets SOC2/HIPAA security standards," including encryption at rest/transit and least-privilege IAM roles. Shadow IT: Unauthorized tools or "rogue" databases used by client employees. This is often where the "cleanest" and most useful data actually lives. System of Record SoR : The authoritative data source for a given piece of information e.g., SAP for finance, Salesforce for CRM . FDEs must identify this to avoid building on "stale" data replicas. Agent2Agent A2A Protocol: An open standard used in the Google Agent Development Kit ADK that allows autonomous agents to discover, hand off tasks, and communicate with one another across systems. Workflow Agents: Deterministic agents in ADK SequentialAgent , ParallelAgent , LoopAgent that follow fixed logic paths rather than relying on an LLM to "plan" the next step. Grounding: The process of connecting an LLM to "Ground Truth" data via RAG or Google Search to ensure its responses are factual and cite-able. AutoSxS Side-by-Side : A GCP-native evaluation method where an "Autorater" LLM compares two model outputs and provides a structured judgment on which is better and why. Faithfulness RAGAS Metric : A measure of how much the answer is derived only from the retrieved context, essentially a "hallucination score". SOW Statement of Work : The legally binding "fence" around your project. If it’s not in the SOW, it’s "Scope Creep". MSA Master Services Agreement : The high-level contract between your company and the client that defines the legal relationship before any specific work starts. Cost of Inaction CoI : A strategic argument used to drive projects forward by quantifying how much money/efficiency the client loses every day they don't deploy the solution. UAT User Acceptance Testing : The "moment of truth" where actual client employees use the tool. If they don't "accept" it, the project isn't done, regardless of how good the code is. Day 2 Operations: Everything that happens after the FDE leaves. Includes monitoring, retraining models, and training the client's internal "Run Team". This repository is curated and maintained by: We built this resource to help bridge the gap for engineers moving into high-stakes, client-facing AI and Data roles. Feel free to connect with us on LinkedIn to discuss FDE, GCP, or Agentic Systems Found a resource that changed your FDE career? - Open a PR. - Ensure the link is high-quality and free or worth the cost . - Add a 1-sentence description of why it matters for FDEs.