Kalpa Connect - fde-academy-lab/c2-content-factory GitHub Wiki
Kalpa Connect
Mobile, broadband and subscription services. The unit where the people who make a number look good have already left, and where almost every metric is a blend of populations that should never have been averaged.
Code
CON. The home of survivorship, cohort mixing, and the gap between predicting something and being able to do anything about it.
Where it sits in the locked ladder
Build 3, Week 9. The stakeholder is Ananya Bose, COO, and her problems are the Week 7 and 8 text problems at telecom scale. Every sub-problem carries an output contract and a cost ceiling, which is what makes them different from the Week 8 exercises:
| # | The sub-problem | The contract on it |
|---|---|---|
| 1 | Churn early warning from support text | An action window that a retention team can actually use |
| 2 | Ticket triage | A fixed schema the output has to match every time |
| 3 | Incident summarisation | A citation contract |
| 4 | Plan recommendation | A refusal rule |
| 5 | Reply drafting | A cost ceiling, and repeatability |
The AI-free observed debug drill runs on the Wednesday morning of that week: each learner debugs a deliberately broken training loop from evidence alone.
Connect is also the last of the three transfer domains from Week 5, usually on churn and support.
1. Who runs it, and what each of them wants
| Role | Measured on | What they want to hear | What they will resist |
|---|---|---|---|
| VP Subscriptions | Blended retention, on a board slide | That retention improved and there is a reason to be proud of | A within-channel view that shows the improvement was a mix shift |
| Head of Acquisition | Signups this month, cost per signup | That the cheap channel is working | Being charged for what happens to those subscribers next year |
| Head of Retention | Saves, and offers accepted | That targeting the highest churn scores works | Being told the highest scorers have already decided |
| Network Operations lead | Incidents, and mean time to restore | That the network is stable | A definition of incident that cannot be reclassified |
| Head of Support | Containment and cost per contact | That the assistant handles ninety percent | A per-issue count that includes the customer's three retries |
The tension that generates most cards here: acquisition is measured monthly and retention over a year, so a quarter can be won by buying subscribers who will churn after the reporting window closes. Every survivorship card in the bank has its natural home in this unit.
2. The data estate
| What exists | The catch it carries |
|---|---|
| Signup date, plan, acquisition channel, cancellation date | Channel mix moves between cohorts, so the blended rate answers a question nobody asked |
| Usage and billing records | Involuntary churn from a failed payment sits in the same column as a decision to leave |
| Support tickets and chat transcripts | The transcripts record what agents said, including what they said wrongly, with no label for which |
| Network incident logs | What counts as an incident is a classification made by the team being measured |
| The plan catalogue, with historical versions | Plans were renamed, so a plan-level trend crosses a rename and looks like a shift in demand |
The absence that defines the unit: there is no record of which promotion ran on which channel. Every before-and-after comparison in Connect is missing the one variable that would explain it, and asking for it is the move.
3. The KPIs, and how each one is gamed
| KPI | What people say it is | The denominator that decides it | How it breaks or gets gamed |
|---|---|---|---|
| Churn rate | Cancellations divided by the base | The base at the start of the period, split voluntary from involuntary | Involuntary churn from failed payments hides inside the same number and has a completely different fix |
| Retention by cohort | Share still subscribed at month three | Within channel, cohort by cohort | Cut a cheap high-churn channel and the blended rate rises while total retained subscribers fall |
| ARPU | Revenue divided by subscribers | Which subscribers, and whether a suspended line counts | Suspending rather than cancelling keeps the numerator and shrinks the denominator |
| Customer acquisition cost | Spend divided by signups | Per channel, within an attribution window | A shorter attribution window makes any channel look better |
| Lifetime value to CAC | The health of the business | A forecast divided by a fact | LTV is a projection presented as a measurement, and its assumption is rarely written down |
| Satisfaction | How customers feel | Respondents, and the response rate | Nine percent response measures the people who still had the energy to answer |
| Containment | Contacts resolved without a human | Per issue, never per session | A customer who retries four times gives you three extra contained sessions |
| Mean time to restore | Incident response | Per incident, with a fixed definition of incident | Reclassify a long incident as three short ones |
The KPI this unit exists to teach: absolute retained subscribers alongside the rate. A rate can improve while the business shrinks, and Connect is where a learner sees that happen with their own hands.
4. Case families, mapped to modules
| Module | Case family here | The teaching version |
|---|---|---|
| 1 Foundations of AI and Data | Cohort retention, blended against within-channel, and the definition change in a release | S-CON-AN-02, the retention that improved because the unhappy left |
| 2 Applied Machine Learning | Churn prediction, the action window, uplift against propensity, plan recommendation | Predicting churn nobody can act on in time, then fixing the label rather than the model |
| 3 Deep Learning and Neural Networks | Usage sequences and anomaly detection on network telemetry | Where a threshold rule is the baseline that has to be beaten |
| 4 Natural Language Processing | Support ticket triage and routing | Class imbalance and a label taxonomy nobody agrees on |
| 5 and 6 Generative AI | Network incident summarisation, a plan-question assistant | An evaluation set built by the people who wrote the knowledge base, which is the trap |
| 7 Production AI | Guardrails, containment measured honestly, and a control that fails open | S-CON-GA-16, where blocks fell to zero for the wrong reason |
| 8 and 9 Agentic AI | Multi-agent escalation between triage, retention and billing | Where handoff between agents becomes the failure surface |
5. The situations hosted here
| Page | Cards |
|---|---|
| Analytics | S-CON-AN-02 (the retention that improved because the unhappy left), S-CON-AN-08, S-CON-AN-09 |
| Machine learning | S-CON-ML-10 to S-CON-ML-12 |
| GenAI and agents | S-CON-GA-14 to S-CON-GA-16 |
6. What Connect must never be used for
- A card whose twist is only arithmetic. Simpson's paradox is a mechanism here, and a card that is nothing more than the paradox restated is a riddle rather than a situation.
- Real telecom data or any operator's published figures presented as Kalpa's. Kalpa's numbers are invented and labelled as invented.
- The first survivorship lesson and the first cohort lesson on the same day. They are the same idea wearing two coats, and a room needs to meet them a week apart.