The Project: the Quillpath Support Desk

Every lesson in this course is set at one company, on one project. Quillpath is made up. All people, workspaces, tickets, invoices, rules, dates and numbers in the project are fictional and simulated. The real cases that open each lesson are separate, cited, and clearly marked as real.

The company

Quillpath sells project-management software online: boards, cards, checklists and automations, used by small teams around the world. It has four plans, Free, Team, Business and Enterprise, priced per member seat. Customers write to support in English, German, Spanish, Japanese and Hindi.

Money in this course is in US dollars, the currency Quillpath bills in and the currency the model providers price in.

The problem

Quillpath's support desk receives about 1,800 tickets a week and has six agents. The median ticket waits 7.5 hours for a first reply. Most tickets are the same few questions: a charge that looks wrong, a locked account, how to export data, whether a refund is due. The answers are already in the help centre.

Maya, the Support Lead, wants an assistant that:

1. Triages every ticket: a category, a priority, and the help-centre article that answers it, if there is one.
2. Drafts a reply from the help centre, citing the article, for an agent to review before it is sent.
3. Says "this needs a person" when no article answers the ticket.
4. Handles refund requests by the written policy, and never moves money without a person approving.
5. Costs at most one cent of model fees per ticket, on average.

You are the engineer building it.

What this course uses the project for

This course is about the model itself: what it is, how it reads tokens, how it chooses words, how to prompt it, how to make it reliable, how to adapt it, how to test it, how to secure it and how to run it. The support desk is the instrument. Every idea is measured on the same tickets, so you see what a technique changes on a task you already know.

It is not the BrightCart store of the AI Engineering course, the Harbourline Freight engagement of the Forward Deployed Engineer course, or the returns desk of the Agentic AI Engineer course. Quillpath is a different company, with a different product, policy and cast.

The people

Customers who recur in examples: Ana Duarte, owner of Ardent Design (workspace W-201), who sends T-1001 about a duplicate charge; Leo Park, a member of the same workspace, who sends T-1034; Grace Mwangi (Hollis Bakery, W-202), Jonas Feldmann (Feldmann Architekten, W-203), Yuki Sato (Studio Nami, W-204) and Ruth Pell (Pell & Rowan, W-205 and W-206). Minor characters may appear in a single lesson. The main cast above does not change.

The rules

Code enforces every rule that moves money or touches an account. The prompt only describes them. The exact wording, the help centre, the 36 labelled tickets, the billing records and the refund rule as code are in 06-shared-data.md.

Quillpath's systems

From Module 6 on, the assistant reaches these systems through five tools with fixed names, so that every module calls them the same way:

The structured outputs also have fixed names. Triage holds category, priority, kb_article (an article id or None) and needs_human. DraftReply holds text, citations (article ids) and needs_human. Lessons may add fields; they do not rename these. Module 6 Lesson 1 adds one optional field to Triage: amount_usd, the amount the ticket names, from 0 to 10,000 USD, or None.

TinyLM and the other models

Hosted models do not let you look inside, so the course trains its own: TinyLM, a GPT-style model with 1,071,872 parameters (a 2,048-slot vocabulary for byte pair encoding tokens (the tokeniser trained on the shared corpus fills 1,134 of them, because every word is already one token), context 128 tokens, width 128, 4 layers, 4 attention heads, tied input and output embeddings). Module 1 builds and trains it in Colab on the corpus from 06-shared-data.md, and later modules open it up to show probabilities, sampling, the KV cache, attention dilution and LoRA. It is useless as an assistant and excellent as an instrument. Never read its output quality as evidence about real models.

The hosted models are the course defaults in 02-setup-colab-and-llm-providers.md. Where a lesson needs a small open model it can download and adapt, it uses SmolLM2-135M-Instruct from Hugging Face (Apache 2.0, checked 9 October 2026).

The canonical offline token counter is tiktoken's o200k_harmony encoding, which tiktoken maps to the gpt-oss models (checked in tiktoken 0.14.0). Provider counts include chat formatting and can differ by a few tokens; Module 2 shows how much.

Dates

• Course date: Friday 9 October 2026. Model names, prices and documentation were checked on this day. Lessons say "(checked October 2026)".
• The project: Monday 1 June 2026 to Friday 11 September 2026, fifteen weeks. Lessons are told in the present tense, as if you are there.
• The data date: Monday 8 June 2026, in week 2. The 36 tickets are a snapshot of that morning's queue. Code measures "days since" from that date (DATA_AS_OF), never from the computer's clock.

Key events

Lessons refer to these events. They are fixed, so that every module tells the same story. Week N starts on the Monday shown. All of them are simulated.

The course does not fix the assistant's accuracy, cost or speed in advance. Those numbers come from running your own code, and they are what your write-up reports.

What you build, module by module

One repository: which module adds which file

Everything you keep goes in one repository, a single Python package called supportdesk/, in a folder called supportdesk-repo (on your Google Drive in Colab, so it survives between sessions). Lessons run in Colab, so their code sits in cells; the labs write each piece to its place in the repository. Every lab from Module 2 on starts by checking that the earlier modules' files are there and writing back any that are missing, so you can start at any module. This table lists what each module's lab writes, checked against the labs on 13 October 2026.

Two things are not in any package file. The pilot incident (PILOT_INCIDENT, from 06-shared-data.md) is a notebook constant: lessons and labs that need it paste it from the shared data, and no module adds it to supportdesk/. Module 1's lab says Module 10 adds it; it does not. Module 8's retry with backoff also stays in lesson cells.

The finished support desk

• Triage of every ticket into typed, validated output
• Grounded draft replies with citations, reviewed by a person
• Refund decisions made by a rule in code, with a person approving every refund
• Tools with argument checks, least privilege and errors the model can act on
• A context builder with token budgets, and memory with provenance
• A measured cost per ticket under the one-cent cap
• An eval suite with confidence intervals, human ratings and a calibrated judge, gating every change
• Defences against prompt injection, measured, and redaction of personal data
• A gateway with fallback across providers, caching, quotas and tracing
• A web service in a container, with a rollback plan
• A write-up and a limitations memo with numbers from your own runs

The one idea

A model that is right 90% of the time is not a product. The product is what you build around it: the test set that tells you the 90%, the rule in code that catches the 10% that moves money, and the person who sees what the code cannot check. The pilot incident in week 11 is what happens when the last of those is overloaded.

Changes from the earlier version (for authors)

The earlier version of this course set the same project at a company called Brightlane. What was kept and what changed:

• Renamed. Brightlane became Quillpath, because "Brightlane" was too close to BrightCart, the store in the AI Engineering course, and both projects build a support assistant. The support lead, Maya, the twelve help-centre article ids, the plans and prices, the refund policy and the supportdesk/ package name are kept.
• Smaller data. 72 tickets became 36 (24 dev, 12 test), so a lesson can paste the slice it needs. The six categories, four priorities, gold article and answerable flag are kept, and the categories now have written definitions.
• TinyLM kept, built in the course. It was a shipped file; the platform has no downloads, so Module 1 now trains it in Colab from the corpus in 06-shared-data.
• Real outputs only. The earlier version showed some hand-written stand-in outputs, labelled illustrative. Every output in this course comes from a real run. A scripted stand-in may still be used to test plumbing (parsers, loops, retries), and is then named as such and never presented as a model's answer.
• Prices fixed. The earlier price table billed Groq's cached input at the full input price. Groq gives cached input a 50% discount; the dated rates are in course-guide/02.
• Tokeniser dropped. An unsourced "claude-legacy" tokeniser was removed. Anthropic's token counting appears only as an official-docs snippet, labelled "not run here".
• New cast. Joel, Anika, Tomasz, Lena and Ravi are named owners for the budget, security, the platform and the labels. The pilot incident is new.