Marlene Castellanos at the Belize Tourism Board has eleven custom data pipelines built by consultants and an IT person who have all departed. The new minister has asked her whether to keep maintaining all of this or switch to managed tools. She is open that she does not know how to evaluate that. You do — because you have built one of the alternatives a managed tool replaces in nearly every prior project.
The discipline skills: a per-pipeline build-vs-buy framework that prices total cost of ownership including staff time, not just hosting; a three-year cost forecast under two scenarios (keep custom vs migrate selectively) with the seasonal pattern modelled; a transition plan that respects the cruise-season reporting calendar; an audit that surfaces the hidden constraints Marlene does not volunteer; a recommendation document calibrated for Marlene to defend to the minister in her own vocabulary.
The AI-direction lesson: the build-vs-buy question is asked at the architecture level across the whole stack, not for one source. AI defaults to "managed everywhere" because that is the defensible-on-paper modernisation story. AI under-prices staff time. AI proposes enterprise governance tooling because that is what ranks in search results. AI proposes a big-bang migration that ignores the cruise season. The discipline is that you design the evaluation framework yourself before any AI direction begins, and AI executes inside it — every recommendation defended by your experience having built the alternative, every cost number including the staff-hours line.
Your Role
You are the architect of the evaluation. You read Marlene's email, reply in her register, open a discovery thread that surfaces the eleven-pipeline inventory and the constraints she does not volunteer (NDA-bound cruise line APIs, paper-form spending surveys), audit against a representative-shape inventory bundle — the audit is what surfaces the 2023 hard-coded currency rates Marlene does not know are there — design the per-pipeline framework, populate it with AI doing the mechanical work inside your criteria, build the cost forecast, stage the transition into the off-season, and hand Marlene a recommendation plus a rebuildable cost-forecast model.
The scaffolding is thin. The brief names the build-vs-buy ask and three constraints in passing — staff capacity, cruise-season reporting cannot be disrupted, the economic impact report must continue as it is. Everything else is your design.
The AI relationship sits at supervisory-to-autonomous inside a framework you have authored. AI does not propose the framework. AI executes the assessment inside it and you verify each output against the criteria you set.
What's New
Last project you designed AI infrastructure from scratch for Domingos Soares at INE Timor-Leste — AGENTS.md alongside CLAUDE.md, skills, hooks, MCP scopes, agent contexts, a Codex CLI portability cross-check, a dual-register handoff, with JT Thompson reviewing the integrated system.
This project no colleague reviews the evaluation — Marlene is the only client check-in. (Yolanda Reyes, a cost and operations consultant, is reachable on demand for a narrow sanity check on whether the staff-time cost of ownership is priced honestly; she volunteers nothing and makes no call for you.) Genuinely new: the build-vs-buy question across an eleven-pipeline portfolio rather than per-source; staff-time TCO as the central economic discipline (the pipeline that costs $50/month to host but $2,000/month in staff time when it breaks twice a quarter); a three-year forecast under two scenarios with a seasonal pattern; a transition plan staged into the off-season; a recommendation written for a non-technical director to defend to a minister; the budget constraint woven through every interaction.
The hard part: holding the per-pipeline discipline when the easy story is "migrate everything to Fivetran." The NDA-bound cruise line APIs cannot route through a third-party vendor without renegotiating data-sharing agreements. The paper-form spending survey has no managed replacement. The commodity visitor-count pipeline is a clean Fivetran candidate. The recommendation is source-by-source, defended by what you have built, and sized to a team that does not include any engineers.
Tools
- Claude Code, BigQuery, dbt, Dagster, Soda Core, GitHub Actions, Git + GitHub,
AGENTS.md+CLAUDE.mdapplied as a short engagement memory — all carry-forward. - Managed-tool references named in the evaluation but not implemented: Fivetran, Airbyte, a managed CDC service, Monte Carlo or SYNQ for observability, and the enterprise governance tools (Collibra, Atlan) that get explicitly ruled out at BTB's scale.
No new tools. What is distinctive is the use: established tools direct the evaluation work, and managed tools are evaluated against what you have built rather than against vendor marketing.
Materials
You'll receive:
- A project governance file (
CLAUDE.md) — the engagement repository's session context: scope, framework-before-execution sequence, verification targets, commit convention. - BTB operational picture (
btb-context.md) — scale numbers, the 50-person agency with the three-people-touch-data team named, the eleven-pipeline portfolio at Marlene's depth of description, the minister's expectation, the staff and consultant history. - Build-vs-buy reference (
build-vs-buy-reference.md) — the literature you cite when designing the framework: managed-ingestion trade-offs, managed CDC vs Debezium, managed observability vs custom Dagster + Soda Core, governance tooling at small-team vs enterprise scale, TCO models including staff time, transition-risk patterns for seasonal businesses. - Discovery prompts (
discovery-prompts.md) — starter questions in Marlene's register for Unit 1: what each pipeline does and costs, how often each breaks, the sharing-agreement question, the spending-survey question, the minister's reporting constraints, the team's capacity in season vs off-season. - Recommendation template (
recommendation-template.md) — the shape of the deliverable only: audience, register, sections. The content is yours. - A pipeline-inventory seed script (
scripts/seed-btb-pipeline-inventory.py) — generates a representative BTB-shaped inventory underreference-data/(eleven pipeline manifests, BigQueryINFORMATION_SCHEMAcost samples, Dagster run-history samples, staff-incident-log samples, NDA and sensitivity flags, hard-coded constants where present). Held in reserve through discovery; invoked at the start of Unit 2.
Deliberately not provided: the evaluation framework, the per-pipeline assessment criteria, the cost-model parameters, the transition-staging rules, the recommendation register. The INE Timor-Leste engagement workspace from last project does not carry forward — fresh project root for BTB.