Fieldpath

Syllabus

Four courses, thirty days

This is the curriculum, not a wrapper around someone else's playlist. Each course is seven to nine full lessons: concepts in detail, a 120-minute session, drills, and checks.

LITDays 1–7 · 14 hours

Speak the language

Become fluent in how modern AI actually works — models, tokens, prompts, and cost — so you can sit in a technical room without translating through someone else.

Literacy

  1. Day 1

    120m

    The map: where a delivery manager actually lands

    Orientation — pick the game before you train for it

    • The jobs are not 'AI engineer'
    • Three layers — do not confuse them
    • How this month is built (60 hours, not a degree)
    • The 2-hour protocol
    • The 2025–26 hiring reality
    • What 'enough technical depth' means
    • Artifacts vs badges

  2. Day 2

    120m

    How a model actually works (without a PhD)

    Literacy — next token, context, and why it hallucinates

    • It is a next-token machine
    • Tokens and the context window
    • Hallucination is a feature of the objective
    • Weights vs context vs tools
    • Temperature, sampling, and why 'creativity' is a setting
    • What to say when a VP asks 'does it know our data?'
    • Four controls that replace 'please don't hallucinate'

  3. Day 3

    120m

    The model landscape: what to use, when, and who owns it

    Literacy — vendors, open weights, and a decision you can defend

    • You are buying an interface, not a soul
    • A working taxonomy, not a leaderboard
    • Data paths decide the vendor more than quality does
    • Five questions that prevent accidental lock-in
    • Azure OpenAI, Bedrock, and Vertex are doors
    • Eval-driven model choice
    • When a smaller/cheaper model wins

  4. Day 4

    120m

    Prompting as a professional skill

    Craft — system, spec, examples, and failure

    • A prompt is a spec with a probabilistic contractor
    • The levers that actually move quality
    • Failure modes you should test on purpose
    • Prompt library as an operating asset
    • System, user, and tool — who is allowed to say what
    • Few-shot, JSON, and prompt versioning
    • How you review a prompt in a 15-minute huddle

  5. Day 5

    120m

    Tokens, cost, and unit economics

    Numbers — the RAID item nobody puts on the RAID log

    • The meter is tokens in, tokens out
    • Caching, batching, and the quiet multipliers
    • Worked envelope: copilot math
    • Worked envelope: RAG math
    • The 10× stress and why demo cost is a lie
    • Put it on the RAID log
    • What you say when finance wants a number tomorrow

  6. Day 6

    120m

    The four interfaces: chat, structured, embeddings, tools

    Architecture — stop saying 'chatbot' for every request

    • Chat is a UI, not an architecture
    • Structured output is how AI enters systems of record
    • Embeddings are for find, not for chat
    • Tools make the model able to act — and able to break things
    • Four interfaces at a glance
    • Chat vs system of record
    • Write-tool risk, named like any other RAID

  7. Day 7 · Lab

    120m

    Lab: a delivery prompt system you can use on Monday

    Build the week-1 artifact

    • The five prompts
    • How you test without a platform
    • Week 1 close
    • Prompt lifecycle
    • Operating note and data rules
    • Scoring a baby eval, then bringing it to Monday
    • The five-prompt artifact in a hiring loop
BLDDays 8–14 · 14 hours

Build enough to be dangerous

Get your hands on the four primitives every AI product is made of: an API call, embeddings, retrieval, and a tool-using loop.

Building

  1. Day 8

    120m

    Python just enough to ship a thin slice

    Build — the 20% of the language that shows up in every demo

    • The only types you need this month
    • Functions are named processes
    • JSON is the wire format of this whole industry
    • Secrets, files, and what 'running it' means
    • What you write vs what an engineer writes
    • A 40-line script is four verbs

  2. Day 9

    120m

    Calling an LLM API like a grown-up

    Build — request, stream, fail, retry, log

    • The request is boring on purpose
    • Streaming is UX, not magic
    • Failures you must name in the design
    • Logs are evidence and a liability
    • Retries, timeouts, and idempotency are the RAID version of a client
    • Temperature, tokens, and the defaults you freeze for production

  3. Day 10

    120m

    Embeddings and semantic search

    Build — the retrieve in retrieval

    • A vector is a point in meaning-space
    • Chunking is a product decision
    • Hybrid search is the grown-up default
    • Retrieval eval: recall is the adult metric
    • Chunk size, overlap, and k are product knobs with costs
    • A 10-question retrieval eval you can defend

  4. Day 11

    120m

    RAG: the enterprise pattern

    Build — retrieve, stuff, generate, cite

    • RAG is a pipeline, not a feature checkbox
    • The generator is a prisoner of the packet
    • Citations are a product requirement
    • When RAG is the wrong hammer
    • Refuse-when-empty is a product feature, not a failed demo
    • Owners, first-release UI, and the one-pager you will reuse

  5. Day 12

    120m

    Why RAG fails in production

    Reality — the failure catalog you will walk into

    • Stale, duplicate, and conflicting truth
    • ACL leaks are a career-class incident
    • Garbage in, fluent garbage out
    • Citation theatre and over-promise
    • The no-go memo and the thinner slice
    • Re-index, ownership, and the RAID you actually run

  6. Day 13

    120m

    Agents and tools: when a loop is worth it

    Build — ReAct, caps, and why 'autonomous' is a smell

    • An agent is a loop, not a vibe
    • Tools are APIs with extra superstition
    • Caps are the product
    • When the loop earns its keep — and when it does not
    • A three-tool spec you can hand an engineer
    • What you take into the lab tomorrow

  7. Day 14 · Lab

    120m

    Lab: design a project-wiki copilot

    Build the week-2 artifact

    • Specificity is the quality bar
    • Architecture of the wiki copilot — boxes you can staff
    • v1 vs non-goals — protect the slice
    • Evals, ACL, cost — the three numbers steering will actually hear
    • RAID, the no-go, and the eight-minute talk track
    • What 'enough building' meant this week
SHPDays 15–21 · 14 hours

Ship like a delivery lead

Apply the job you already have — RAID, vendors, SLAs, governance, change — to AI workstreams that fail in new ways.

Delivery

  1. Day 15

    120m

    Data, PII, and the governance that actually blocks you

    Ship — the workstream before the model workstream

    • Every AI feature is a new place data lives
    • Classification drives the path — not the model brand
    • Data-ready is a definition of done, not a folder that exists
    • DPA, retention, and the contract you actually need
    • PII and secrets in the prompt are incidents waiting for a name
    • Prompt injection is a content problem, not a patch
    • The security conversation you want to have

  2. Day 16

    120m

    Evaluation: how you know it works

    Ship — acceptance criteria for a probabilistic system

    • Three layers, three owners
    • The golden set is a managed artifact
    • Metrics you can defend: groundedness, recall, task success
    • LLM-as-judge is a junior reviewer
    • The eval loop: change, measure, decide
    • Release bar and waivers

  3. Day 17

    120m

    Guardrails and human-in-the-loop

    Ship — the control plane around a model that can be wrong

    • Layers, because any one layer fails
    • HITL is a dial, not a religion
    • The roster is the control
    • When HITL is theatre
    • The incident page: detect, contain, communicate, learn
    • Maker-checker, SOD, and the controls you already run

  4. Day 18

    120m

    Production: latency, reliability, and the ops you already know

    Ship — SLOs for a system that shrugs

    • The request path, with fallback as part of the happy path
    • SLOs that survive nondeterminism
    • Latency, nondeterminism, and what you tell a sponsor
    • Model versions are releases
    • Deprecation is a project, not an email
    • Dashboards a delivery lead should demand

  5. Day 19

    120m

    Build vs buy, vendors, and the POC trap

    Ship — procurement with a spine

    • Buy vs build for a thin slice, not for a platform fantasy
    • The five questions, production edition
    • The POC that cannot accidentally productize
    • Demo-ware, lock-in, and subprocessors
    • Who owns day 91
    • Platform-for-everything and other ways to delay the first slice

  6. Day 20

    120m

    The AI delivery playbook (your actual job)

    Ship — discovery, scoring, RAID, change, value

    • Discovery is jobs, not models
    • Score like a portfolio manager
    • RAID, rewritten for models
    • Change is trust plus a roster
    • Sequence and kill criteria, published before you start
    • HIPPO, portfolio, and saying no without theatre

  7. Day 21 · Lab

    120m

    Lab: the use-case charter and eval plan

    Build the week-3 artifact

    • Four readers, four paragraphs — what steering-ready means
    • Assembly: you already wrote the parts
    • The headings that have to be there
    • The 8-minute talk track
    • Kill criteria, waivers, and what you will not pretend
    • Week 3 close: you can run the workstream
LNDDays 22–30 · 18 hours

Land the role

Pick the AI seat that fits a delivery manager, build the portfolio that proves it, and practice the interviews that hire for it.

Career

  1. Day 22

    120m

    Pick the seat: a real target, not a vibe

    Land — one primary role, one backup

    • Match the last three years, not the fantasy
    • The default pair: delivery plus solutions
    • Read posts like a delivery lead
    • How to read noisy titles
    • Gaps that are allowed vs gaps that fail a loop
    • Ten posts as a fit lab, not a lottery
    • Positioning v1 is a bet, not a biography

  2. Day 23

    120m

    Portfolio: the demo story

    Land — problem, slice, measure, expand

    • The arc they remember
    • The six boxes you can draw from memory
    • Honesty if you did not ship code
    • A script you can steal and then make yours
    • Failure modes you should name before they ask
    • Who gets this story, and who gets tomorrow's

  3. Day 24

    120m

    Portfolio: the delivery artifacts

    Land — charter, RAID, eval, cost on one table

    • The operating pack is a table, not a novella
    • Five-minute steering, not a status novella
    • Pushback A: 'just demo it'
    • Pushback B: 'we need a platform'
    • Pushback C: 'legal won't allow'
    • Two stories, one use case — pick in ninety seconds

  4. Day 25

    120m

    Positioning, resume, LinkedIn

    Land — the paper that gets the loop, not the paper that lists tools

    • The top third does all the work
    • Honest AI bullets — steal these shapes
    • Honest artifact phrasing
    • Keywords with a spine
    • LinkedIn About is positioning v1, not a manifesto
    • The honesty pass as a procedure

  5. Day 26

    120m

    Interview loops: what they actually run

    Land — screens, cases, technicals, take-homes

    • Loops by seat
    • The 90-second intro is positioning, spoken
    • The five-minute messy-program story
    • Technical screens if you are not 'an engineer'
    • Take-homes: a policy, not a vibe
    • Questions you ask them

  6. Day 27

    120m

    Scenario drills: the chatbot-by-Friday

    Land — live judgment under a bad request

    • The spine, every time
    • Drill A — chatbot of all company knowledge by Friday for the CEO
    • Drill B — agent that emails customers when SLA slips
    • Drill C — we already bought Vendor X, stand it up
    • Drill D — legal said no cloud models, now what
    • How to drill so the spine sticks

  7. Day 28

    120m

    Capstone assembly

    Land — one folder, six artifacts, no orphans

    • The pack
    • The six artifacts, and what each must prove
    • The hole protocol
    • Folder flow: how a stranger moves through it
    • The 12-minute outline (not the script)
    • What you tick, and what you send

  8. Day 29

    120m

    Walkthrough: 12 minutes, then questions

    Land — the performance

    • The 12-minute flow, then Q&A
    • The likely eight
    • Voice — short sentences, numbers, owners
    • Run 1, cut 20 percent, Run 2
    • Handling Q&A without becoming a different person
    • Lock it, and do not reopen the month

  9. Day 30

    120m

    The 90-day plan and the job operating system

    Land — after the month, the machine that gets the offer

    • 90 days in the seat (they will ask)
    • The job search as a workstream
    • One next skill, four weeks, not a new stack
    • What you can prove as of tonight
    • The 90-day plan as a sendable page
    • Close the month like a go-live