0/35FDE path lessons understood. Go to your dashboard.

 

Topic 01 of 10

Models & context

How a model reads, counts and stops, and how to make a single call trustworthy enough to put in front of a bank's staff.

  • 10 lessons
  • 5.9 h reading
  • 9.3 h hands-on
  • 3 on the FDE path
Start lesson 01
Progress0 / 10

This track's client

N

Head of Branch Operations, Northvale Bank

Mon 09:12

AI helper for branch staff: can we have something by Friday?

Our branch staff lose a lot of time looking things up in the policy manuals. We want an assistant they can ask “who signs off a cash deposit above the limit?” and get the answer straight away. It goes into the CRM case view, so the Salesforce team says it must be proper structured data, not a chat blob.

  • Latencyp95 ≤ 4 s
  • Input≤ 6,000 tokens
  • Cost≤ $0.01 / answer
  • OutputAnswerV1 JSON
Wednesday0 / 5 lessons

The brief changes on Wednesday. Mark 5 lessons understood to see how.

Suggested order, not the numbered one

  1. Step 1

    Mechanics

    01

  2. Step 2

    One trustworthy call

    050608

  3. Step 3

    Fit the budget

    07

  4. Step 4

    Pick on evidence

    02

  5. Step 5

    When to think harder

    0304

  6. Step 6

    The Wednesday change

    10

  7. Step 7

    Ship it

    09

Lessons

  1. 01Tokens, context windows and why answers go wrongPath · stage 2Read a usage block and a stop reason, and explain the failure before the customer asks. Pro
  2. 02Choosing a model you can defendGates, evidence and an exit plan, so “why this model?” has a five-part answer. Pro
  3. 03Reasoning models and what thinking costsSet reasoning effort per route from eval evidence, not from the default. Pro
  4. 04ReAct, chain of thought and reflection in practiceThe patterns behind every agent framework, and what each one multiplies. Pro
  5. 05System prompts that survive a regulated rolloutNumbered, owned clauses that compliance can read and you can test. Pro
  6. 06Typed output: schemas, validation and tool callsPath · stage 2The CRM wants JSON, every time, with a safe fallback when it isn't. Pro
  7. 07Context on a budget: what to keep, compress and cutFit the evidence, the history and the instructions inside a token and cost ceiling. Pro
  8. 08Answers with receipts: grounding and citationsPath · stage 2Every claim traced to a source, checked in code, or the assistant says it can't. Pro
  9. 09Treat prompts like releasesVersions, changelogs and staged rollout, so a prompt edit never surprises production. Pro
  10. 10Beyond English: language, script and region LabHindi and Hinglish cost more tokens and fail differently. Budget and test per language. Pro

You're done with this track when you can

0/8