"From Vibe Coding to the AI Software Factory" — Ben Rogojan (SeattleDataGuy)

Why this is in the vault

Rogojan lays out a 5-level maturity ladder for how companies actually adopt AI coding tools (Garage → Macro Workbench → Workshop → Assembly Line → Value Factory), which is a clean, citable external articulation of the "harness beats raw model access" thesis RDCO already operates on.

Mapping against Ray Data Co

Rogojan's Level 3 ("Workshop") diagnosis — companies buy Claude Code or Copilot, get disappointing results, then bolt on AGENTS.md files and shared prompts as incremental patches with "a missing flow to it all" — is a plain-language restatement of the harness-engineering-book thesis already anchoring RDCO's positioning: model access alone doesn't produce repeatable output, the harness (state machines, specs, tokenomics tracking, error/recall tracking) is the actual product. His Level 5 ("Value Factory") caveat — a fast, error-free assembly line that produces widgets no one asked for is still worthless — is the same L4→L5 gap RDCO is itself navigating (agent capability without a demand-generation engine attached). Useful as a plainspoken maturity-model vocabulary to borrow when pitching phData or CAF/OI stakeholders on why "just give people Claude seats" undersells the actual problem.

Related

• [[2026-05-26-seattle-data-guy-ai-consultants-utility-thesis]] — same author's adjacent thesis that AI's power without simplicity is why labs build consulting/SI ecosystems; this issue's "Workshop" stage is the mechanism behind that demand.
• [[2026-07-06-seattle-data-guy-fde-dilution]] — same author on forward-deployed-engineering title inflation; both pieces describe companies groping for repeatable AI-delivery structure without one yet.
• [[project_l5_north_star_strategic_direction]] — RDCO's own L4→L5 framing (agent capability vs. demand generation) maps almost one-to-one onto this issue's Assembly-Line-vs-Value-Factory distinction.

⚠️ Sponsorship

Two distinct promotional elements, both disclosed in-text: (1) Estuary — top-of-article pre-amble naming Rogojan as a platform advisor/client-user, the recurring SDG pattern per prior backfill notes; (2) self-consulting CTA — closing P.S. pitches Rogojan's own firm, Codestrap, for "software factory" consulting engagements, meaning the entire maturity-model framework doubles as sales collateral for the service he's pitching in the same breath. Read the 5-level ladder as a persuasive frame ("you're stuck at Level 3, we sell Level 4") as well as an analytical one.

Curation section

• "How to Escape the Ad-Hoc Data Request Trap" — recap of an SDG-hosted webinar with Celina Wong (ex-Ulta/Poppins data leader) on moving data teams from reactive ticket-answering to strategic partnership. Self-cross-promo (SDG's own webinar series), not third-party.
• "Beyond Prediction: Solving the Multiple Knapsack Problem at Scale" — third-party engineering writeup on Uber's Tarot targeting-orchestrator platform, framing incentive allocation across millions of users as a constrained optimization (Multiple Knapsack Problem) rather than a pure prediction task. Genuine third-party link, no visible SDG relationship.