Innovation Train

We started from the client's research corpus and a set of lo-fi mockups, and produced the artefacts a build could actually run on: a written product requirements document, a primary persona, a route map, and an explicit build-versus-buy position that kept the team from rebuilding commodity infrastructure.

The persona became the tie-breaker for every later decision: a Head of People & Organisational Development who buys the assessment, distributes it, and then has to stand up and explain it.

The brand had to signal research credibility to a C-suite audience without borrowing the visual habits of consumer software. We wrote the direction down before drawing it: measured and declarative, closer to an engineering consultancy than a startup. The reference point was a respected mechanic looking under the hood, not a SaaS pitch, and named anti-references kept the work honest.

Colour carries diagnosis rather than decoration: moss marks what is healthy, rust marks what is constrained, brass carries emphasis. A serif display face against a plain grotesque gives the editorial weight of a research publication instead of a product page.

The report is the product. The starting point was a research document assembled in a slide tool: plotting-library charts, default type, findings buried below the fold. We rebuilt it as an eighteen-page system: cover, disclaimer, executive summary, the model, overall profile, benchmark comparison, department comparison, one section per dimension, strengths to protect, gaps, highest-leverage focus areas, a phased action plan, and a closing.

It renders as HTML at slide dimensions so the same source produces a projected deck and a forwarded PDF, with a department-scoped variant for organizations that behave less like one train and more like a railway network.

Every report is written per client, which means a language model drafts the prose. The governing decision was to separate the arithmetic from the argument completely: scoring is deterministic code, interpretation is the model, and the two meet only at composition time.

Deterministic scoring turns survey responses into means, 0–100 scores, benchmark gaps, per-item agreement rates, department scores and flags for the strongest and weakest dimension. The model never sees a raw response and never originates a number; it reads the computed profile and writes what it means. Composition then maps both into a slide-scoped source document, one entry per page.

That separation is what makes the output defensible. Every field in the report carries the layer that produced it, so a reader can always tell a measurement from an interpretation.

The customer-facing application covers the full commercial path: marketing, authentication, plan selection and checkout, assessment setup, live response-rate monitoring, analysis and the report library. All of it sits behind route guards, a billing gate and a content layer keyed by section, so copy can change without touching components.