Independent teaching notes. Not affiliated with xAI or Grok.
Forward Deployed Engineering · self-paced studio
Learn Forward Deployed Engineering by taking an AI system from discovery to production
Discover the workflow, design the system, integrate the real source of truth, keep authority in code, evaluate behavior, deploy, operate, and decide if it created value.
Six weeks on one fictional shop, Northline Orders. Order 1042 is shipped with carrier Parcel North. Order 1043 is paid and has not shipped. Order 1044 is cancelled. Each row has an email you must not leak.
Easy is everyday language. Start there. Medium is for IT. Pro is for people who already understand language models. A week takes about 2–3 hours if you run the lab. There is no live class, no fee, and no certificate. Reading these pages does not credential you.
Northline Orders and the five case-study shops are fictional practice. They are not real clients. Portfolio policy.
One loop, a step budget, a context budget, short-term versus long-term memory, and a first eval that does not need a judge model. Before the loop: a discovery note, one success metric, a pilot boundary, and a written no.
Side effects, a plan that is allowed to be wrong, and a typed handoff so a second agent cannot widen its own authority. Also a field map onto a system that is not three clean rows.
Upsert without rebuilding the world, tombstones, access control in the query, hybrid rank, and input and output checks. Also the pages you will not index, named with the person who owns them.
Traces as test oracles, a judge model you do not fully trust, regression in CI, and a refund gateway that ignores injected instructions. Also their security constraints, each tied to a control or a refusal.
An authenticated HTTP edge, a job table, a container that runs the checks, and a pipeline that does not call a paid model on every keystroke. Also a runbook their on-call can follow without you.
Traces across model calls and tools, token and error signals, cache and rate limits, then a production review of the same order flow. Also an incident note a person sends, and an adoption count that is not page views.
Each page is one fictional engagement, from the first conversation through a production decision or a deliberate stop. They are fictional practice, not real clients and not placements. Northline Orders stays the lab. These five shops are separate stories. Say "I studied this fictional engagement." Do not invent a customer or a metric. See the portfolio policy.
Harbor Lane Clinic. A regional clinic inbox that drafts replies, must not send them, and must not read billing.
Keel and Bin Parts. A parts warehouse where the nightly export is not the live stock count.
Mill City Permits. A permits desk that answers from a named corpus and refuses anything outside it.
Ridge Route Field. A field-service pilot whose only offered login is an admin account, so the write tool stays off.
Lumen Desk. A support team that keeps a parallel run after a green health check, and writes a five-line note when the new path is wrong.
Interview practice for forward-deployed work
A forward-deployed engineer builds the pilot and sits with the customer. A forward operator runs the live pilot, watches failures, and writes the handover. These are role shapes to practice. They are not a certificate and not a job offer from this site.
FDE and forward operator interview practice. A day in the work, questions with short model answers, resume lines that stay fictional, a 30-minute mock, and what not to claim.
Interview kit
A separate page from the practice above. It paraphrases public posts where someone described a forward-deployed or AI engineer interview, and it points each verified round at a fictional case on this site. It is not a script, and it is not part of the interview practice page.
Extra notes for Pro. They do not replace the six weeks. Easy and Medium do not depend on them. No account, no certificate, and no placement.
Prompt, retrieval, or fine-tune. One leave-policy question, three options, the failure each still accepts. The default is do not fine-tune.
Keel and Bin data reconciliation. Required practice for Week 2, at every depth. A messy extract, the live table, SQL for the gap, and a field map that is wrong on purpose.
Identity and audit, added on week 3. Group, row filter, who is logged, and what must never enter the prompt. Harbor Lane billing and chart text stay out by a snippet, not by a policy sentence.
Cost sheet, required on week 6. Runs per day, tokens per run, the cache hit you are assuming, and the call you cut first. You fill the practice numbers. They are not savings.
These notes were written for this studio by its author, Murari Bhartia, as original teaching notes. They are not a rewrite of a vendor course. Grok11 is independent notes, not a school, and not affiliated with xAI or Grok.
Full disclaimer. Every engagement on this site is fictional. Northline Orders is the lab shop. Harbor Lane Clinic, Keel and Bin Parts, Mill City Permits, Ridge Route Field, and Lumen Desk are fictional case studies. None is a real client, and studying them is not a placement. Say "I studied this fictional engagement." Do not invent a customer, a metric, a savings figure, or a job outcome. A passing python3 test_all.py is a lab check, not a credential. This site offers no certificate and no score. The same rules live on the portfolio policy page.
What you will have at the end
Not a platform. A small system you can explain: one request comes in, auth runs, a loop calls an allowlisted tool, retrieval filters by group before the model sees text, a gateway refuses a refund the prompt tried to authorize, a trace line exists without the shopper's email, and a test fails if any of those controls move. That is the production shape. Week 6 uses the same hostile refund. The lab test checks the pieces side by side. It is not one HTTP call that runs every loop. The weeks also ask for practice documents: a discovery note, a field map, a list of pages you will not index, a constraint list, a runbook, an incident draft, and an adoption table. They describe the fictional shop. They are not a certificate, and they are not evidence of a real customer's deployment. The lab does not grade those documents. It also does not grade a bad judgment inside them, such as counting logins as adoption, indexing a whole folder, or calling a deflection rate success.
Practice pages
Single-topic pages pulled out of the weeks, each with the lab code it describes and the case it matches.
ACL before rank. Filter who may read a chunk before you score it. Week 3.
Tombstones in retrieval. Stop an old version answering without rebuilding the index. Week 3.
Reciprocal rank fusion. Keyword plus vector rank, the arithmetic, and the tie-break. Week 3.
Title each page "Northline practice, fictional." If you mention this studio, say you studied the notes. Do not say you were placed on a customer site. A passing python3 test_all.py is a lab check, not a credential. There is no score on these pages, and no certificate. Every rule in one place: portfolio policy.
Week 1. Who does the work, the metric in both wordings, the pilot end date, the out-list, and one refusal.
Week 2. Field map: their name, your name, the source of truth, and what you return when the sides disagree.
Week 3. Corpus list: the owner, the pages in the pilot index, and the pages out. Pay stays out.
Week 4. Their constraints, each with a control or a tool you will not ship.
Week 5. Runbook: down, stop, previous image tag, who may run it, and the pilot account that is not your login.
Week 6. A five-line incident draft, how long traces live, and an adoption table that can say stop. The old screen stays up during a parallel run.
Mind map: terms to recall
Recall the whole studio: three orders, a loop you stop, a gate on writes, groups from the session, checks on actions rather than tone, no secrets in the image, a trace id on the reply, and no certificate or fee.
What you need before you start
Python 3.11 or newer. You do not need a web framework. The checks use only Python's built-in tools.
A terminal, the text window where you type commands. The check command is python3 test_all.py. On Windows it is py test_all.py.
A text editor, so you can read and change the lab files.
Easy still starts from everyday basics: files, a web page, and the idea that a program can answer in sentences. The Python command above is part of the studio, not a surprise later in the week.
Lab files, in one zip
Download lab.zip. Python 3.11 or newer. From the folder where you saved the zip:
unzip lab.zip
cd lab
python3 test_all.py
Expect 31 passes, with no network and no secret key. On Windows the command is py test_all.py. The count is the number that command prints. Build your own version first if you want the practice. The reference is a worked answer you can compare with, not something to paste into a job without reading. A passing run is a lab check, not a credential.
If python3 test_all.py fails, use the checks in the FAQ before you write in.
Why week 2 is in this studio
The notes are original. They are not a vendor course, and they are not a rewrite of anyone else's syllabus. Week 2 is a bridge this studio wrote so the rules for each tool exist before search. The outline these notes started from had no separate week there, so this is not someone else's week 2.
How to move through it
Every week is one lesson at three depths. The address does not change. On the week page, the Easy, Medium, and Pro buttons only show or hide sections. They do not load a different course. The shop, the three orders, and the rules do not change.
Read the week. Leave it on Easy unless the sentences feel too slow, then switch to Medium or Pro. On a week page, the pictures sit above the Easy, Medium, and Pro buttons.
Do the "what to build" list, or open the matching file from the zip and explain each choice out loud.
Run the checks. A passing run uses a stand-in, which is fixed practice code, not a real provider. It only proves the steps you wrote. Week 4 says what you still owe a real provider.
Where the field is unsettled, the text says so. Long-context behavior, judge models, and prompt injection do not have a settled fix. Treat confident blog posts about those three as hypotheses.
If python3 test_all.py fails
There is no class forum. These are the first checks. A question after that can go to info@grok11.com.
You are in the lab/ folder that contains test_all.py and the northline/ folder. Running the file from another directory makes the imports fail.
The command is python3 test_all.py. You need Python 3.11 or newer. python3 --version should report that. python --version on some computers is not Python 3.
On Windows the command is py test_all.py, not python3 test_all.py. py --version should report Python 3.11 or newer.
You do not need a network connection, an account, or a secret key for this command. NORTHLINE_TOKEN is only for the optional local server in the README, not for test_all.py.
Read the first failed check name in the output. The week page names the rule that check locks. A failure means that rule moved, not that the studio is down.
Do not paste order emails, tokens, or the full traceback into a public page. If you still cannot tell which rule failed, email info@grok11.com with the check name and the Python version.