Intencv

The real problem wasn't writing résumés. It was applying without knowing whether your CV actually matched the role.

TL;DR

Time to ship

1 week

Idea to live product

Ownership

SOLO

Product, design & engineering

Systems shipped

4

One connected product ecosystem

Composite score

+45 pts

35% → 80% on one application

Intencv — Beat ATS landing hero with résumé upload flow

+45%

Match-score lift

35 → 80 composite · documented Revolut Product Designer application

4

Systems shipped

Core product, LinkedIn tool, SEO engine, distribution pipeline

20+

Indexed SEO routes

~30 lines of config per page

Most résumé tools generate first and ask questions later. Intencv checks a CV against the actual job posting before AI runs and before payment. The repeatable score becomes the rewrite target, the reason someone can judge value, and the constraint that shaped scoring, paywall, generation, and PDF delivery.

Summary

Client
Intencv — score-first résumé tailoring. Solo founder product, live at intencv.com (2025–2026).
Problem
People paid to rewrite CVs without knowing if they fit the role. ATS noise mattered less than applying blind. No honest read on match before time, money, and reputation were on the line.
What I did
Designed and shipped the full loop: deterministic match score against the live posting, constrained AI rewrite toward that target, Stripe payments, server-rendered PDF delivery, programmatic SEO, and distribution tooling. Not a demo, a paid product. I owned product, UI, and engineering solo — no external design system; type, layout, and interaction patterns are mine end to end.
Business impact
On one documented Revolut Product Designer application, composite match moved 35% → 80% (+45 points) after tailoring. That output is measured product behavior, not a hiring result.
Intencv — match score, signals, and tailored résumé preview in one surface
Upload to match score. Parsing and job URL on the left; deterministic score, signals, and projected lift on the right.

01

What most tools assume

Typical flow: generate, then show a percentage that doesn’t steer anything. When I scored the same CV with an LLM three times, I got 67%, 71%, and 64%. Fine for a vibe check, useless as a reason to pay.

The score couldn’t be decorative. If it wobbled, it couldn’t justify payment or tell the rewriter what to fix. A stable, pre-generation read against the live posting gave both: a paywall users could evaluate and a target the model had to hit.

01

What most tools assume

One documented Revolut Product Designer application

Generic CV composite

35%

Tailored output composite

80%

+45pts

BeforeAfter
  • Keywords2871
  • Skills4585
  • Relevance3578
  • Context2862

Directional pillar movement; composite total is the shipped system output.

Composite before
35 percent
Composite after
80 percent
Gain
45 percentage points
Keywords
28 to 71
Skills
45 to 85
Relevance
35 to 78
Context
28 to 62
Intencv — upload modal with parsing state and job URL field
Intencv — match score panel with composite readout and pillar signals
Same session, two panels. Résumé upload and job link on the left; deterministic composite score and pillar signals on the right, before anyone pays.
Intencv — tailored résumé ready state with match score, signals, and tracked edits
Score-first output. Match readout, rewrite diff on the résumé, and ready-to-apply actions in one surface.

02

Key decisions

Decision matrix
  • Score before paywall

    Why
    users see a concrete reason to pay first. Some people with low scores leave; I don’t treat every drop-off as healthy, but charging before a number is worse.
  • Deterministic scoring over LLM scoring

    Why
    repeatability matters when payment hangs on the read
    Trade-off
    rule-based matching can miss semantic equivalents a model might catch.
  • No account to start

    Why
    upload and score before signup. Less friction up front; more abuse surface to handle later.
  • Merge AI output onto the base CV

    Why
    a bad suggestion can’t replace the whole document
    Trade-off
    less freedom to restructure the original layout.
  • Server-rendered PDF via Browserless

    Why
    the file is the product; client render varies by device. ~2–4s latency and higher per-use render cost, traded for consistent output.

Score before paywall — users see a concrete reason to pay first. Some people with low scores leave; I don’t treat every drop-off as healthy, but charging before a number is worse.

Deterministic scoring over LLM scoring — repeatability matters when payment hangs on the read. Trade-off: rule-based matching can miss semantic equivalents a model might catch.

No account to start — upload and score before signup. Less friction up front; more abuse surface to handle later.

Merge AI output onto the base CV — a bad suggestion can’t replace the whole document. Trade-off: less freedom to restructure the original layout.

Server-rendered PDF via Browserless — the file is the product; client render varies by device. ~2–4s latency and higher per-use render cost, traded for consistent output.

Intencv — exit-intent modal offering one free tailored CV before leave
Paywall sequencing. Exit-intent capture: one free CV, no card, single-session expiry.

03

What shipped

03

What shipped

Score-first pipeline. Deterministic scoring runs before any LLM rewrite or payment.

  1. 1

    CV + job URL
  2. 2

    Protected job-content fetch
  3. 3

    Deterministic 4-pillar score (free, no LLM)
  4. 4

    Optional learning context
  5. 5

    AI rewrite constrained to score target
  6. 6

    Merge onto original CV
  7. 7

    Server-rendered PDF
  8. 8

    Score validation & optional stretch pass

CV and job URL, SSRF-protected fetch, free deterministic 4-pillar score, optional learning context, constrained gpt-4o-mini rewrite, merge onto the original CV, Browserless PDF, then score validation and optional stretch pass.

  1. CV + job URL
  2. Protected job-content fetch
  3. Deterministic 4-pillar score (free, no LLM)
  4. Optional learning context
  5. AI rewrite constrained to score target
  6. Merge onto original CV
  7. Server-rendered PDF
  8. Score validation & optional stretch pass
CV parsing fallback
  1. 01

    Client extraction
  2. 02

    Server dual-library retry when layout corruption is detected
  3. 03

    Vision OCR when text extraction fails

Real CVs break clean extraction. Client parse first, server dual-library retry when layout corrupts, vision OCR when text fails. Roughly 10% of uploads I saw needed fallback for coordinate layouts. Observed sample, not an industry statistic.

Systems on one stack
  • Core product
  • LinkedIn post canvas
  • Programmatic SEO engine
  • Reddit / Quora distribution pipeline
  • 20+ indexed SEO routes
  • ~30 lines of config per page

Beyond the core product: admin LinkedIn post canvas (/linked), programmatic SEO template (20+ routes, ~30 lines of config per page), and a Reddit/Quora distribution pipeline (scrape, dedupe, rules, AI quality gate, human review, no auto-posting).

Stack: React, Node.js, OpenAI, Stripe, Redis, Vercel. Entry tier $9/application. Margin driven more by PDF rendering than model tokens.

Intencv — LinkedIn resume transformation canvas, dark theme
Distribution canvas. Before/after résumé transformation for LinkedIn; dark and light export frames.

Intencv — LinkedIn resume transformation canvas, dark theme. Alternates with: Intencv — LinkedIn resume transformation canvas, light theme

LinkedIn feed — Intencv post with before/after canvas and match lift in the wild
In the feed. Same export frame in a real LinkedIn post: copy, Revolut role context, and the generated before/after graphic.

Proof

  • +45%

    Composite match on one documented Revolut Product Designer application. 35% on the generic CV, 80% after tailored output (+45 points). Single application, not a cohort average.

  • ~65%

    Approximate funnel: uploads that reach a viewed score before pay (~65%). Observed shape, not proof that the score lift drove conversion.

  • 4

    Systems shipped on one stack: core product, LinkedIn admin canvas, programmatic SEO, distribution pipeline with human review.

04

Documented outcomes

Observed approximate funnel
CV uploaded~100%
Score viewed~65%
Paid~20%

Observed product funnel — not evidence that match-score lift caused conversion or hiring outcomes.

Systems shipped: four products on one stack (core, LinkedIn tool, SEO engine, distribution pipeline); 20+ indexed SEO routes from one template (~30 lines of config per page).

Documented Revolut Product Designer application in Proof above. Tailored output on file, not a hiring result.

Observed funnel (approximate): ~100% CV uploaded → ~65% score viewed → ~20% paid. Shapes how people move through the product; not evidence that the score lift caused hires.

Systems shipped: four products on one stack (core, LinkedIn tool, SEO engine, distribution pipeline); 20+ indexed SEO routes from one template (~30 lines of config per page).

Evidence

01

Score before paywall

Free deterministic 4-pillar score against the live job URL. Generation and payment only after the user sees an auditable read on fit.

Intencv — deterministic match score against the live job posting before paywall

02

Deterministic over LLM scoring

Same CV returned 67%, 71%, and 64% across three LLM scoring calls. Unusable as a paywall signal. Pillar math is stable run to run.

Intencv — stable deterministic pillar scoring instead of variable LLM percentages

03

Deliverable is the PDF

Server-rendered PDF via Browserless. Client-side render varies by device; the résumé file is the product outcome, not the UI chrome.

Intencv — server-rendered PDF résumé as the product deliverable

What I’d repeat

  • Build the trust mechanism before you optimize the conversion flow — if the number isn’t credible pre-pay, nothing downstream fixes it.
  • Use AI where it earns its cost; use repeatable rules for decisions that have to look the same every time (especially anything tied to payment).
  • Design the system, not just the screen — scoring, paywall, generation, PDF delivery, and unit economics are one product decision, not separate features.