The Project: the Tidewell Cycles Inbox

Every lesson in this course is set at one company, on one project. Tidewell Cycles is made up. The shop, its people, its customers, its products, every message and every number in the project are fictional and simulated. The real cases that open each lesson are separate, cited, and clearly marked as real.

The company

Tidewell Cycles is an independent bike shop and repair workshop in Bristol with a small web shop. It sells city bikes, e-bikes, a gravel bike, a children's bike, accessories and workshop services. Customers write in by email, through the web form on the website, and in product reviews. All of it lands in one shared inbox.

On Wednesday 1 July 2026 Tidewell started delivering to Germany, Spain and France. Until then almost every message was in English. From July about four in ten are in German, Spanish or French.

Money in this course is in pounds sterling (GBP), the currency Tidewell sells in. Hosted model prices are in US dollars, as the providers publish them; Module 10 says which is which wherever both appear.

The problem

The inbox received 200 to 270 messages a month in the spring, and about 430 a month since July (simulated, measured from the shared data). One person, Ines, reads them part-time. She forwards repair bookings to Dev in the workshop, product questions to Callum, and the rest she answers herself; spam stays with her too, by default, because nothing filters it. Spam, a battery fault and a summer sale have all arrived in the same queue, and nobody can say which problems are getting worse.

Rosa, the owner, wants five jobs done:

1. Route every message to the right person (intent classification).
2. Pull out the order number, product, fault and preferred date, so nobody retypes them (extraction).
3. Tell her each Monday what customers are unhappy about, and whether it is getting worse (sentiment and topics over time).
4. Suggest the right help-centre article to Ines while she replies (semantic search and question answering).
5. Handle German, Spanish and French messages without hiring for each language (multilingual NLP and translation).

You are the contractor Rosa hires to do it. Module 1 Lesson 1 sets the five jobs; the capstone delivers them.

What this course uses the project for

This course teaches natural language processing from the ground up: text as data, rules, tokens, linguistic structure, vectors, classifiers, evaluation, sentiment, topics, embeddings, transformers, extraction, question answering, summarisation, translation, and large language models used as one more tool. The inbox is the instrument. The same hand-written test set (hard_test, 80 messages) is scored by every method as it arrives: the rule router in Module 1, TF-IDF models in Module 4, embeddings in Module 6, a fine-tuned DistilBERT in Module 7 and hosted LLMs in Module 10. By the end you have one honest comparison table, with cost.

It is not the BrightCart store (AI Engineering), the Harbourline Freight engagement (Forward Deployed Engineer), the returns desk (Agentic AI Engineer), the Quillpath support desk (Large Language Models), Hollin Answers (RAG), the notes assistant (MCP), the TinyFormer lab (Transformers) or the Booknook shelf (Python). Tidewell is a different company, with a different cast, data and problem.

The people

Customers come from a fixed list of 60 fictional names in the shared data (30 women and 30 men, 36 writing in English and 8 each in German, Spanish and French), deliberately varied in origin so that the Module 10 fairness check has something to measure. No customer is a recurring character except Mrs Joan Hartley of Clifton, whose long, polite, increasingly angry thread about her Volt 2 battery (TH01) is the running example for summarisation and escalation. Do not add named staff; a customer may appear in one lesson only.

The products

Fourteen products, fixed. SKUs, names and prices come from PRODUCTS in 06-shared-data.md and never vary.

Identifiers have fixed shapes, so that Module 1 can find them with regular expressions:

The rules

The help centre (KB01 to KB24 in the shared data) is the written policy. The facts lessons rely on most:

Two project rules are not in the help centre:

• Labels come from Ines's written definitions, LABELS in the shared data. Seven intents: order_status, returns_refund, repair_booking, product_question, warranty_fault, praise, spam. Three sentiments: neg, neu, pos. When a message fits two intents, the label is the one that needs action first, and hard_test records the second.
• Amara's rule for hosted models (agreed with Rosa in Module 10): only redacted text, with names, emails, phone numbers and postcodes replaced, may be sent to a hosted model, and the lesson that does it names the provider's own data terms rather than guessing them. This is the project's rule, not legal advice.

The data

All Tidewell data is simulated and generated by code with a fixed seed, random.Random(20260831), in plain Python, so every learner gets identical data. 06-shared-data.md holds the code and the fingerprints to check it.

Template messages flatter template-learnable models: a classifier trained and tested on MESSAGES scores far higher than it will on real mail. That is why HARD_TEST is the headline number, and Module 4 measures the gap rather than claiming it.

Dates

• Course date: Wednesday 14 October 2026. Library versions, model names and prices were checked on this day or on the date shown beside them.
• The data window: Monday 2 March to Sunday 30 August 2026, 26 weeks.
• DATA_AS_OF = Monday 31 August 2026, the morning the inbox was exported. Code counts days to DATA_AS_OF, never to the computer's clock.
• The project: Monday 7 September to Monday 16 November 2026. Module n is project week n. Lessons are told in the present tense, as if you are there.

Key events in the data

Fixed, so that every module tells the same story. All simulated.

The project weeks

Fixed names

So that modules fit together, these names do not change:

• The repository is tidewell-inbox/; the package is inbox/; tests live in tests/.
• inbox/data.py holds the shared data blocks from course-guide/06, copied exactly.
• Every intent model, from the rule router to the LLM classifier, offers predict(texts: list[str]) -> list[str], returning one of the seven intents (or "abstain" once Module 5 adds abstaining). The bake-off calls every model the same way.
• inbox/evaluate.py offers report(y_true, y_pred), which every module from 5 on uses to score a model on HARD_TEST.
• inbox/llm.py holds the setup cell from course-guide/02 and ask(model, messages), created once.

One repository: which module adds which file

In Colab each lesson writes the files it needs with %%writefile and imports them; on a laptop the same files form the repository. Tests run with !pytest -q in Colab and pytest -q locally. Every lab from Module 2 on starts by checking that the earlier modules' files are present and writing back any that are missing, so you can start at any module.

The finished inbox service

• Clean, normalised text, with personal data redacted before storage
• Order numbers, products, dates, money and faults extracted, validated and scored field by field
• Every message routed to Ines, Dev or Callum, in four languages, with spam set aside in a spam folder and an "abstain" (sent to Ines) for what fits nowhere
• One comparison table: rules, TF-IDF, embeddings, DistilBERT and two hosted LLMs on the same 80 hand-written messages, per class and per language, with cost and speed
• A hybrid router that sends only the hard cases to a hosted model
• Help-centre search and an answer finder for Ines, scored with recall@5, MRR and exact match
• Rosa's Monday report: what customers are unhappy about, and whether it is getting worse
• A name-swap fairness check, a model card and a list of what broke

The one idea

Measure the simple method first. A regular expression, a word count and a logistic regression are cheap, fast and explainable, and on many inbox jobs they are hard to beat. Every model in this course has to earn its place against them on the same 80 hand-written messages, and the capstone keeps the cheapest method that does each job well enough.