HAMMAD YOUSUF

FREE RESOURCE · FROM THE "ROADMAP" REEL

Forward Deployed Engineer: a free 6-week roadmap (2026)

By Hammad Yousuf · 18 October 2026

A six-week skills sprint for developers moving into Forward Deployed Engineer (FDE) work. Each week has one free official resource and one milestone, and you keep extending one project: a document chatbot in week 3, an agent with tools in week 4, a scored test suite in week 5 and a live service in week 6.

What a Forward Deployed Engineer does

  • Palantir: its Forward Deployed Software Engineers work directly with customers, often on site, and Palantir compares the job to a startup CTO's. Fortune describes other tech companies as copying Palantir's playbook. (Palantir posting, Fortune)
  • OpenAI: its Forward Deployed Engineering team works with customers to turn research into production systems and embeds closely with the people who use them. (OpenAI posting)
  • Google Cloud: its GenAI FDE is described as an embedded builder who takes frontier AI products to production inside customers. (Google posting)

In short: AI engineering plus customer deployment. You build with LLMs, and you make the result work on the customer's real data, systems and constraints.

Why now (numbers, sources, dates)

  • FDE job postings rose more than 1,000% in January-August 2026 compared with the same months of 2025, and more than 4,600% compared with 2023. Source: Lightcast data reported by Fortune, 3 Sep 2026.
  • The median advertised FDE salary was more than $188,000, against roughly $145,000 for traditional software engineers. Source: the same Lightcast data in the same Fortune article. Advertised pay varies a lot by country and level.
  • India: FDE postings were up 1,807% year on year in June 2026, nearly 16 times higher in July and more than 6 times higher in August. Source: Indeed data reported by People Matters/PTI, 15 Oct 2026.
  • These are counts of job postings, not hires.

Who is hiring (official job boards, 18 Oct 2026)

  • Palantir: 77 of its 310 open postings have "Forward Deployed" in the title, including new-grad tracks and a Forward Deployed Software Engineer role in Abu Dhabi. (Palantir jobs)
  • Google: 52 FDE-titled roles on the first five pages of Google Careers results, across Google Cloud (Generative AI), Google DeepMind and Google Public Sector. (Google Careers search)
  • OpenAI: 26 FDE postings, including Forward Deployed Engineer - UAE in Abu Dhabi, posted 21 Sep 2026. (OpenAI jobs)
  • Anthropic: 6 FDE postings in Paris, London, Munich, New York, San Francisco and Seattle. (Anthropic jobs)

In the UAE: two Abu Dhabi FDE roles were open on 18 Oct 2026, at Palantir and OpenAI.

Honest expectations (read this first)

  • Six weeks builds the skills and one portfolio project. It does not make you hireable from zero.
  • Google's FDE III (Generative AI, Google Cloud) asks for 5 years of software development in Python or a similar language. Across Google's FDE listings, the minimum ranged from 1 to 8 years on 18 Oct 2026. (posting, search)
  • OpenAI's UAE FDE role asks for 5+ years of engineering or technical deployment experience that includes customer-facing work. (posting)
  • Palantir is the exception: it runs new-grad Forward Deployed tracks. (Palantir jobs)
  • Our pacing assumption (not from a source): about 8-10 hours a week. If you are new to Python, give week 1 two weeks.

What FDE postings actually ask for (18 Oct 2026)

  • Python or a similar language: a minimum qualification in Google's FDE postings. OpenAI's UAE role asks for production code in Python, JavaScript or comparable stacks. Week 1.
  • Building with LLMs: OpenAI's UAE role asks for systems built or deployed with LLMs or generative models. Week 2.
  • RAG with vector databases: a minimum qualification for Google's FDE III (data pipelines using vector databases and RAG-like architectures). Week 3.
  • Agent frameworks: Google's FDE III prefers experience with multi-agent frameworks such as LangGraph, CrewAI or ADK. Anthropic's FDE postings in Paris, London and Munich list agent development. Week 4.
  • Evals: 18 of OpenAI's 26 FDE postings mention evals, and Anthropic's Paris, London and Munich FDE postings list evaluation frameworks. Week 5.
  • Shipping on a cloud, with a design you can defend: Google's FDE III asks for experience architecting AI systems on cloud platforms such as GCP and taking AI solutions from concept to launch for customers. Anthropic lists deployment at scale. Week 6.
  • Customer work: Google's FDE III asks for experience leading technical discovery sessions with customers, and Palantir's FDSEs work directly with customers. Practise it with the discovery-call prompt below.

Sources: Google FDE III, OpenAI FDE - UAE, OpenAI jobs, Anthropic jobs, Palantir posting.

Day 0: free setup

  1. Install Python 3.15, the current release on python.org on 18 Oct 2026: python.org/downloads
  2. Install an editor and Git: VS Code, Git
  3. Get a Gemini API key in Google AI Studio (Google sign-in needed): aistudio.google.com/apikey. Keep it in a .env file, never in code or chat.
  4. Know the free tier: Gemini 3.8 Flash input and output are listed as "Free of charge" on the free tier, and your free-tier rate limits are shown inside AI Studio. Paid Gemini 3.8 Flash costs $0.75 input / $3.75 output per 1M tokens through 31 Dec 2026, then $1.50 / $7.50 from 1 Jan 2027. (pricing, rate limits)
  5. Privacy caveat: Google's pricing page says free-tier prompts are used to improve its products. Use public data or your own documents, never client data, on the free tier.
  6. The UAE and India are on the Gemini API's available regions list.
  7. Optional, fully local models: Ollama.

The project you will build

Pick one problem that looks like real customer work, for example: "answer staff questions from a company's policy PDFs, and look up live data when the documents are not enough". Keep one repo the whole way: a Python script (week 1), a model call (week 2), a document chatbot (week 3), an agent with tools (week 4), a scored test suite (week 5), and a live service with a design doc (week 6).

Week 01 of 06

Week 1: Python

Goal: small Python scripts that talk to web services and keep going when a request fails.

Week 1 milestone: API script that survives failures

Prompt for your coding agent (Claude Code, Codex or similar):

prompt 1 of 9
In a new folder fde-project, set up a Python virtual environment and write fetch.py: it calls the public JSON API at [API URL], retries up to 3 times with exponential backoff on timeouts and 5xx errors, logs every failure to errors.log, checks that the fields [field list] exist, and saves clean records to data.json. Add pytest tests that fake a timeout and a 500 response. Explain each try/except block in one line.

Week 02 of 06

Week 2: LLM APIs

Goal: call a model from your own code, and explain what happens inside it.

Week 2 milestone: first model call on the free tier

prompt 2 of 9
Following https://ai.google.dev/gemini-api/docs/quickstart, add chat.py to fde-project: a command-line chat that calls gemini-3.8-flash with my free-tier key from a .env file (never hard-code it), keeps the conversation history, prints the token count of each reply if the SDK returns it, and exits cleanly on Ctrl+C. Then quiz me with 5 questions on tokens, context windows and transformers and correct my answers.

Week 03 of 06

Week 3: RAG

Goal: a bot that looks up the answer in your PDFs and shows which passage it used.

  • Free official resource: LangChain: Build a semantic search engine. LangChain is MIT-licensed; version 1.4.6 was current on PyPI on 18 Oct 2026.
  • Next step: LangChain: RAG with Deep Agents
  • Embeddings: Gemini Embedding 2 text input is free on the free tier (pricing).
  • Managed alternative: Gemini File Search. Storage and query-time embeddings are free, indexing embeddings are billed, and the free tier allows a 1 GB store.
  • Optional video: freeCodeCamp, Learn RAG From Scratch (2 h 33 min, April 2024). It was recorded before LangChain 1.x, so some of its code is dated: YouTube
  • Milestone: a chatbot over 10+ of your own PDFs that cites the chunk it answered from, and says so when the answer is not in the documents.

Week 3 milestone: RAG chatbot that cites its source

prompt 3 of 9
Following https://docs.langchain.com/oss/python/langchain/knowledge-base, build rag.py in fde-project: load the PDFs in ./docs, split them, embed them with a Gemini embedding model using my free-tier key from .env, store them in a local vector store, and answer questions with gemini-3.8-flash. Every answer must name the file and page of the chunk it used, and reply 'not in the documents' when nothing relevant is retrieved. Check the current LangChain docs for package names before installing anything.

Week 04 of 06

Week 4: ADK agents

Goal: give the chatbot tools so it can act, not only answer.

Week 4 milestone: ADK agent with two tools

prompt 4 of 9
Install google-adk and, following https://adk.dev/get-started/python/ and https://adk.dev/tools-custom/function-tools/, wrap my week-3 retriever as a function tool called search_docs and add a second tool that calls [a real API]. Give each tool a clear docstring so the model knows when to use it. Run the agent locally the way the get-started guide shows and show me one conversation where it uses both tools.

Week 05 of 06

Week 5: Evals

Goal: prove the agent works with test cases and scores.

  • Free official resource: ADK: Why evaluate agents and the evaluation criteria
  • Evaluation is built into ADK: adk eval runs .test.json test files and evalsets. The default criteria are tool_trajectory_avg_score (exact match of the tool-call trajectory, threshold 1.0) and response_match_score (ROUGE-1 similarity to a reference answer, threshold 0.8). LLM-judged criteria are also available; they use model calls.
  • Read: Hamel Husain, AI evals FAQ
  • Tool: Arize Phoenix, free to self-host under the Elastic License 2.0
  • Optional video: Laurie Voss (Arize), hands-on evals workshop for agents (2 h 04 min): YouTube
  • Milestone: 20+ test cases scored on both criteria, plus a before/after table for one fix.

Week 5 milestone: 20 evals and a before/after table

prompt 5 of 9
Using https://adk.dev/evaluate/, write 20 test cases for my agent in ADK's .test.json format: 10 normal questions, 5 that need the API tool and 5 that the documents cannot answer. Include the expected tool calls and a reference answer for each. Run adk eval with tool_trajectory_avg_score and response_match_score, show the failures in a table and propose one fix. After the fix, rerun and give me a before/after table.

Week 06 of 06

Week 6: Ship it

Goal: a live URL, a design you can defend, and a demo a customer can follow.

  • Free official resource: ADK: Deploy to Cloud Run (the command starts with adk deploy cloud_run)
  • Cost: the Cloud Run free tier with request-based billing covers 180,000 vCPU-seconds, 360,000 GiB-seconds and 2 million requests a month, counted per billing account, and a billing account is required (Cloud Run pricing). New Google Cloud customers get $300 in free credit (Google Cloud free). Set a budget before you deploy.
  • System design: System Design Primer (free, CC BY 4.0)
  • Milestone: a live endpoint, a one-page architecture doc (data flow, scaling, cost, failure modes) and a 2-minute demo recorded as if for a customer.

Week 6 milestone: deploy to Cloud Run + one-page design

prompt 6 of 9
Deploy my ADK agent to Cloud Run following https://adk.dev/deploy/cloud-run/ (adk deploy cloud_run). Before deploying, explain what the Cloud Run free tier covers on https://cloud.google.com/run/pricing and help me set a billing budget. Then write a one-page architecture doc: data flow, where secrets live, how it scales, estimated monthly cost at 100 and 10,000 requests a day, failure modes and what I would monitor. Finish with a 2-minute demo script for a non-technical customer.

Talk about your project

Use these four points in every application and interview:

  1. The problem, in the customer's words.
  2. Who the customer was and what data you had.
  3. The metric you moved (your week-5 before/after table).
  4. What you would build next, and what it would cost to run.

More copy-paste prompts

Build my personal 6-week FDE plan (Claude or ChatGPT)

prompt 7 of 9
You are my study coach for a 6-week Forward Deployed Engineer (FDE) skills sprint. My background: [current role, languages I know, years of experience]. Time I can give: [hours per week]. Target employers: [e.g. Palantir, Google Cloud, OpenAI, Anthropic]. Use this structure and keep ONE project growing the whole way: Week 1 Python glue code (APIs, JSON, error handling); Week 2 LLM APIs on the Gemini API free tier; Week 3 RAG over my own documents with citations; Week 4 an agent with tools in Google ADK; Week 5 evals with adk eval; Week 6 deploy to Cloud Run and write a one-page system design. For each week give me: one free official resource, a day-by-day task list that fits my hours, the milestone that proves I finished, and what to skip if I fall behind. Be honest about which weeks I can compress or must stretch given my background, and end with 3 questions an FDE interviewer would ask me about this project.

Pick my free stack

prompt 8 of 9
Help me pick the stack for a portfolio project: an agent that answers questions from a company's documents and calls one real API. Constraints: free tiers only, I am in [country], and the data is [public / my own / a client's]. Compare: the Gemini API free tier vs a local model with Ollama; a LangChain vector store vs Gemini File Search for retrieval; Google ADK for the agent; Cloud Run for hosting. For each choice give the free-tier cost, setup time and the data-privacy trade-off (the Gemini API free tier uses prompts to improve Google's products, so client data needs the paid tier or a local model). Check the current official docs and pricing pages before you answer, cite the URL for every price or limit, and finish with one recommended stack and a 5-step setup checklist.

Practise the customer side: a discovery call

prompt 9 of 9
Act as a customer at a [industry] company with a messy problem: [e.g. the support team is drowning in policy questions]. I am the Forward Deployed Engineer on a 30-minute discovery call. Stay in character, give vague answers at first, and only share the key constraints (data access, security rules, the success metric) when I ask good questions. After 10 exchanges, step out of character and score my discovery: what I asked well, what I missed, and how I should scope a 2-week pilot.

Where to apply

Sources (all checked 18 Oct 2026)

Every fact on this page was checked against the source next to it. Job boards, prices and docs change often, so treat each source as the final word.

  • FDE job postings up more than 1,000% in Jan-Aug 2026 vs the same period of 2025 and more than 4,600% vs 2023; median advertised FDE salary more than $188,000 vs roughly $145,000 for traditional software engineers (Lightcast data, Fortune, 3 Sep 2026); other tech companies copying Palantir's playbook. Source ↗
  • India FDE postings up 1,807% year on year in June 2026, nearly 16x in July, more than 6x in August (Indeed data, People Matters/PTI, 15 Oct 2026). Source ↗
  • Palantir: 77 of 310 open postings have "Forward Deployed" in the title, incl. new-grad tracks and an Abu Dhabi FDSE role (re-checked 18 Oct 2026). Source ↗
  • Palantir FDSEs work directly with customers, often on site; responsibilities compared to a startup CTO; Abu Dhabi location. Source ↗
  • Google: 52 FDE-titled postings on the first 5 Careers results pages; minimums range from 1 to 8 years across FDE listings (18 Oct 2026). Source ↗
  • Google FDE III (Generative AI): embedded builder; 5 years software development in Python or similar; architecting AI systems on cloud; leading technical discovery sessions with customers; RAG with vector databases; ADK among preferred multi-agent frameworks. Source ↗
  • OpenAI: 26 FDE postings, 18 of them mention evals (re-checked 18 Oct 2026). Source ↗
  • OpenAI Forward Deployed Engineer - UAE (Abu Dhabi, posted 21 Sep 2026): partners with customers to turn research into production; 5+ years incl. customer-facing work; Python, JavaScript or comparable; LLM systems. Source ↗
  • Anthropic: 6 FDE postings (Paris, London, Munich, NYC/SF/Seattle); Paris, London and Munich list agent development, evaluation frameworks and deployment at scale (re-checked 18 Oct 2026). Source ↗
  • Python 3.15.0 is the current release (docs title 3.15.0 Documentation). Source ↗
  • Python tutorial section 8.3 Handling Exceptions (Week 1 resource and recording). Source ↗
  • Gemini 3.8 Flash free tier input/output "Free of charge"; paid $0.75 in / $3.75 out per 1M tokens through 31 Dec 2026, then $1.50 / $7.50 from 1 Jan 2027; free tier "Used to improve our products: Yes"; Gemini Embedding 2 text free on free tier. Source ↗
  • Gemini API quickstart default model gemini-3.8-flash (Week 2 resource and recording). Source ↗
  • Free-tier rate limits are shown in AI Studio. Source ↗
  • UAE and India are supported Gemini API regions. Source ↗
  • Gemini File Search: storage and query-time embeddings free, indexing embeddings billed, 1 GB store on the free tier. Source ↗
  • LangChain 1.4.6, MIT licence (PyPI, 18 Oct 2026). Source ↗
  • LangChain semantic search tutorial: index documents, add_documents (Week 3 resource and recording). Source ↗
  • ADK is Google's open-source Agent Development Kit (Apache-2.0), optimised for Gemini but model-agnostic, available in Python, TypeScript, Go, Java and Kotlin. Source ↗
  • google-adk 2.12.0, requires Python >= 3.11 (PyPI, 18 Oct 2026). Source ↗
  • ADK docs live at adk.dev (moved from google.github.io/adk-docs). Source ↗
  • ADK function tools (Week 4 resource and recording). Source ↗
  • ADK evaluation: adk eval, .test.json files and evalsets; tool_trajectory_avg_score (exact match of tool-call trajectory) and response_match_score (ROUGE-1) criteria (Week 5 resource and recording). Source ↗
  • ADK default criteria thresholds: tool_trajectory_avg_score 1.0, response_match_score 0.8; LLM-judged criteria. Source ↗
  • Deploy an ADK agent with adk deploy cloud_run (Week 6 resource and recording). Source ↗
  • Cloud Run free tier (request-based billing): 180,000 vCPU-seconds, 360,000 GiB-seconds, 2 million requests per month per billing account; billing account required. Source ↗
  • $300 in free credit for new Google Cloud customers. Source ↗
  • System Design Primer is free (CC BY 4.0); "How to approach a system design interview question" (Week 6 recording). Source ↗
  • Arize Phoenix is free to self-host under the Elastic License 2.0. Source ↗
  • Optional videos and lengths: freeCodeCamp Learn Python 4:26:52. Source ↗
  • Andrej Karpathy, Let's build GPT, 1:56:20. Source ↗
  • freeCodeCamp Learn RAG From Scratch, 2:33:11, April 2024 (pre-LangChain 1.x). Source ↗
  • ADK Masterclass, 3:12:48, May 2025 (pre-ADK 2.x). Source ↗
  • Laurie Voss (Arize), hands-on evals workshop for agents, 2:04:18. Source ↗
  • Google ML Crash Course LLM module (Week 2 resource). Source ↗
  • Hugging Face LLM Course chapter 1 (Week 2 resource). Source ↗
  • Anthropic, Building effective agents (Week 4 reading). Source ↗
  • Hamel Husain, AI evals FAQ (Week 5 reading). Source ↗
  • Google AI Studio API key page (needs Google sign-in). Source ↗

Counts and prices were checked on 18 Oct 2026 and change often; Gemini 3.8 Flash paid prices rise on 1 Jan 2027. This is a learning plan, not a job guarantee. The reel that links here uses AI-generated avatar stills of Hammad.

About the video: the reel that sent you here is presented by an AI avatar of Hammad Yousuf.

Keep going