Learn by Mentoring, a leadership-skills gym built for Tangent

Tangent is an AI talent network for job seekers from low-income backgrounds. Managers at member companies mentor those job seekers, and Tangent had noticed the mentors were developing as leaders too. Learn by Mentoring turns that into a product: a leadership-skills gym where a manager learns one skill, rehearses it safely and sees their progress. Sales enablement is the first vertical because its return is the easiest to measure.

The brief was a one-page proposition from Tangent and a technical specification we wrote together, then iterated with Joe Adams, Tangent's co-founder, in the open rather than through wireframes. Tangent created the Google Cloud project and covered the running costs. We covered the engineering, and everything we built belongs to Tangent.

Sign-in leads into a short tour that points at each part of the product, then a baseline: a goal, development priorities, an industry track and a self-reported score on each of six competencies, tapped on a pill scale. Saving locks it. Everything that happens afterwards is measured against that starting point, so the progress page can show an honest comparison rather than a number that drifts.

The six competencies and their eighteen observable behaviours are Tangent's own framework. We encoded it directly, chapter by chapter, rather than generating curriculum content with a model.

Ten core chapters, then a four-chapter track in tech sales or customer success. Each chapter is a slide deck: a cover, a narrated video lesson, the idea, a framework, a worked example and a per-question knowledge check. Narration is generated once per scene with ElevenLabs and cached. If the provider is unavailable the lesson still plays, with browser narration and captions.

Chapters open in any order. An earlier version gated them, and the gates kept drifting out of sync across screens, so the ordering became a suggestion and the product got simpler.

Every chapter has a scenario. The mentor reads a brief about the jobseeker they are about to coach, consents to the recording and holds a real-time voice conversation with an AI mentee who plays that person, in a British voice matched to the character. An audio-reactive aura shows who is speaking, and the transcript builds as they talk.

On hang-up the transcript goes to Gemini with the scenario and the chapter's competency rubric. The result comes back as structured JSON, is checked for shape and range, and is saved as a debrief: what worked, what to sharpen and a score per behaviour.

The debrief is the product's honest moment. It says what worked, what to sharpen next time and one thing to try, then scores each of the chapter's three behaviours out of five with the line from the transcript that earned it. A pass bar of 3.5 decides whether the chapter is done. The rehearsal below was run for this page, with the mentor's side spoken from a script; the transcript and the scores are what came back.

Each competency is the mean of its three behaviours, using the latest assessment of each. A rehearsal moves a score because Gemini scored the transcript against the chapter's rubric. Finishing a chapter or passing a quiz never does. Scores can fall as well as rise, and the progress page keeps the locked baseline beside the current value so the change is visible.

Sales and L&D leads see sign-ups, training completion, competency coverage across the cohort and example AI feedback, and can download a standalone manager report that prints to PDF.

The curriculum studio lets them edit any chapter's title, competency, lesson, worked example and quiz from a shared template, create new chapters, preview them and export the catalogue.

Three route handlers sit beside the app: one swaps the ElevenLabs key for a short-lived signed URL, one runs the evaluation, one generates narration. The API on Cloud Run in Tangent's Google Cloud project answers only a same-origin proxy that signs each request as the frontend's own service account, with a GET-only allowlist. The app degrades on purpose: no provider audio means browser narration and captions, and no cloud means bundled data and a badge that says so.

For the four-week demo, product state is a reducer persisted in the browser, with every action mapped to a named endpoint on the Cloud Run service, so moving it server-side is one swap point rather than a rewrite. The Postgres schema, seed data, container and build pipeline are in the repository.

Every week ended with a production preview. By handover the product had seven unit tests over the journey rules and an end-to-end suite that walks first login, the manager report, the mobile layout, chapter completion and authoring.

If you need a product designed and built quickly and handed over properly, including where AI belongs in it and where a person should stay in the loop, contact us.