Comeback is a voice-first AI career agent for workers displaced by AI. It does the three things that most determine whether a displaced worker lands well — in the order the job hunt actually needs them:

  1. Pivot Map ($19) — an honest occupation outlook, a transferable-skill graph, 3–4 target roles with live demand and salary data, and a gap plan built on evidence projects instead of course-collecting. Built from Google-Search-grounded Gemini research, a schema-constrained report, and an editorial report page.

  2. The Gym — job-specific voice mock interviews generated from the real job description and company, rubric-scored every session (private coaching feedback, never employer-facing), with a dossier and score history. Two modes: full-duplex live voice (Gemini Live — talk naturally, interrupt, auto-ends after about 5 questions and scores the session) and push-to-talk.

  3. The closed loop — gym scores and interview outcomes re-weight where the agent spends your effort (warm paths vs. applications, across role targets), and every reallocation is written to a visible decision log that explains why it shifted your effort this week.

First wedge: customer-support and call-center professionals displaced by AI support bots — the largest, most identifiable AI-displaced cohort, voice-skilled by profession, with a clean pivot into CX ops, AI-bot supervision/QA, and conversation design. The narrative writes itself: the people AI bots displaced become the people who run the bots.

Pricing: free Reality Check (no card) leads in; the full Pivot Map is $19 one-time; membership is $29/mo for unlimited Voice Gym (incl. live voice), pipeline, Connection Engine, and the weekly allocation agent.

Three independent ranking rounds over 480 candidate ideas converged on voice-first AI career tools. A needs-first review of the displaced job hunter's journey then reshaped the product: the two needs that most determine outcomes — direction (is my occupation shrinking? where do my skills transfer?) and relationships (warm paths beat cold applications) — lead the product, and the application machine is demoted to workhorse. The wedge chose itself: support professionals are the most concentrated, most public AI-displacement cohort of 2025–26, and they're voice-skilled — a perfect fit for a voice-first product.

  • Gemini Live (gemini-3.1-flash-live-preview) for voice sessions — gated on a week-1 latency test that passed at median time-to-first-audio 727ms, making full-duplex voice UX viable. The live-voice Gym is finished: it captures both sides of the conversation, auto-ends after about 5 questions, and scores the session.
  • Gemini for pivot mapping, job-description analysis, and rubric scoring — Google-Search-grounded research (gemini-3.5-flash) feeding a schema-constrained report model (gemini-3.1-pro-preview).
  • Google Cloud Run hosts the app; Firestore (native) stores orders, leads, scores, and the decision log; Cloud Scheduler runs the weekly effort-allocation agent.
  • Stripe hosted Checkout for payments — card data never touches the server; a signed webhook marks paid and triggers generation, with a success-page fallback so a slow webhook can't strand a customer.
  • Every Gemini call is logged (op, model, tokens, latency) to Cloud Logging — this doubles as the contest's required production evidence. Admin-token-gated endpoints expose the funnel and allocation evidence.- Validating that voice latency was good enough for a real-time UX before committing the whole product to it (the 727ms gate).
  • Mapping a candidate-side coaching product cleanly outside employer-side automated-employment-decision rules (e.g., NYC LL144) — solved by keeping rubric scores private to the candidate, never employer-facing.
  • Flat pricing only: success/placement fees risk state employment-agency licensing, so the business is software plus coaching, never a placement agency.
  • Local gcloud broken by corporate TLS interception — drove all Google Cloud config/secret/domain ops through a Node + ADC REST driver instead.
  • Getting full-duplex Live voice working through an ephemeral browser token: the token must lock the full session config (model, modalities, transcription), and the browser connects with an empty config — otherwise the session closes immediately.

  • Live in production at https://comeback.careers on Cloud Run, with Firestore, Gemini, and Stripe LIVE.

  • The full paid funnel proven end-to-end (live Stripe charge to an 80-second Gemini Pivot Map to Firestore "ready"), validated with a founder self-test that was then refunded.

  • Feature-complete across Pivot Map, Voice Gym (full-duplex live voice plus push-to-talk, both rubric-scored, with dossier / history / chart), pipeline, connections, and the weekly allocation agent — 394 tests passing.

  • An 11-function AI operating model: internal operations run autonomously and log every decision; outbound actions are agent-drafted and human-sent by design (the consent moat).

  • Compliance-by-design from day one: candidate-side only, consent-first data, deletable voice recordings, never used for training.

  • Live in production at https://comeback.careers on Cloud Run, with Firestore, Gemini, and Stripe LIVE.

  • The full paid funnel proven end-to-end (live Stripe charge to an 80-second Gemini Pivot Map to Firestore "ready"), validated with a founder self-test that was then refunded.

  • Feature-complete across Pivot Map, Voice Gym (full-duplex live voice plus push-to-talk, both rubric-scored, with dossier / history / chart), pipeline, connections, and the weekly allocation agent — 394 tests passing.

  • An 11-function AI operating model: internal operations run autonomously and log every decision; outbound actions are agent-drafted and human-sent by design (the consent moat).

  • Compliance-by-design from day one: candidate-side only, consent-first data, deletable voice recordings, never used for training.

  • Full-duplex voice on Gemini Live is viable for a coaching product when you measure first and design the UX to the measured latency.

  • For a revenue-judged contest, evidence-generating work (revenue, agent logs, testimonials) has to ship early and rough; polish ships last.

  • The hard part of a career product is regulatory positioning, not the AI — getting candidate-side-only and flat-pricing right is what keeps it shippable.

  • A high first-purchase price suppresses the first sale for a cash-strapped audience; lowering the entry to $19 (tripwire) removes that barrier while subscriptions carry the economics.

  • Distribution to the support-pro wedge (LinkedIn build-in-public plus support-professional communities) to earn arms-length revenue across June–August.

  • A first-month credit and an annual plan to deepen the subscription bias the rules reward.

  • Connection Engine (consent-based warm-path detection plus call rehearsal), a draft-and-one-tap-approve search agent, and an outplacement B2B pilot or LOI as trajectory evidence.

Built With

Share this project:

Updates