Video summary

I'm Building an App to Make $10,000/Month - Episode 2

Main summary

Key takeaways

Business

Business / strategy summary

  • Three app ideas evaluated (choose the “best” business option):

    • AI content creation for creators (faster)
    • Beautiful AI portrait generator
    • Explicit content generator (for “gooners”)

  • Decision constraint (non-coding): he chose the portrait app partly because it avoids personal risk/social friction—e.g., “wouldn’t require explaining browser history to my wife.”

  • Core consumer-product thesis:

    • Growth = distribution + marketing + funnel that gets payment.
    • Retention = build something users return to.
    • But his immediate goal is narrower: make $10,000/month, even if users don’t come back (short-term monetization > long-term retention for MVP experiments).

MVP / go-to-market approach (playbook)

  • “Fake it to validate demand” (aligned with YC-style advice):

    • Build a landing page and sell before the full product exists.
    • Use a manual/illusion-based backend behind a “loading bar” to mimic instant results.
    • Rationale: avoid spending months building something nobody wants.

  • Landing page funnel design (intended: ad → trust → pay):

    • Page positioning: “AI portraits for your dating profile”
    • Trigger: payment link appears when users are “most vulnerable” (timing/psychology focus)
    • Waitlist gating: “Due to high volume, we’re very selective”

Concrete execution steps taken

  • Created portfolio content without a finished product:

    • Since there were no photos of real customers, he generated “fake customers” by transforming existing portraits via AI.
    • This unblocked the ability to launch marketing assets and test conversion.

  • Deployed the web app (operations stack):

    • App built with Node.js
    • Hosted on Hostinger for cost-effective hosting of multiple Node.js apps (up to five on business shared hosting)
    • Deployment workflow:
      • Connect to GitHub
      • Configure environment variables
      • Create a MySQL database
    • Experiment-ready plan: he intends to build/host multiple apps during the series.

Product approach (how he plans to generate likeness)

  • Tried closed models (discarded):

    • Google/OpenAI approaches: don’t train on your face → portraits don’t match the subject.

  • Adopted open-model + LoRA strategy (framework/process):

    • LoRA (fine-tuning add-on) trained on a person to shift outputs toward likeness.
    • Example concept: train LoRA on an art style (e.g., children’s art) or on a specific celebrity (e.g., Billie Eilish).
    • Constraint: LoRAs work with open models, so he chose that route.

  • Base model choice:

    • Selected Z Image Turbo after evaluating open-model options (his “evaluation process” was reviewing a recent Austris AI YouTube video).

  • Aspirational-yet-accurate pipeline (manual “taste” step he monetizes):

    • Find a professional, well-lit celebrity/model photo with aesthetic similarity.
    • Morph the celebrity into the customer using AI, but dial the blend (slider between aspirational and accurate).

  • Key realization affecting product design:

    • Even with the “same parameters,” output quality depends heavily on the random seed.
    • He doesn’t want to fully automate generation yet because customers aren’t paying for raw inference—they’re paying for his selection/taste to pick the best result out of many.

Customer validation & pricing experiments (metrics / KPIs)

  • Revenue target / time horizon framing:

    • Primary KPI: $10,000/month (goal for the whole venture).

  • Pricing experiment intent (no specific final numbers):

    • He wants to do A/B tests on pricing after gaining interest.
    • Later, he asks users whether they’d pay $35 for outputs.

  • Test outcome (qualitative “conversion intent” signals):

    • Interview feedback from testers:
      • One says no to paying $35 (and implies these apps aren’t their typical behavior).
      • Another says “absolutely not” to “generate like five photos,” but asks if they’d get just one.
      • Multiple reactions:
        • Some like aesthetics (glow, clean look)
        • Others flag uncanny valley or “too sexy,” “too skinny,” or not looking enough like them.

Actionable recommendation emerging from the episode

  • Positioning mismatch risk: the biggest barrier isn’t image quality—it’s whether the user feels the output matches their face and whether they perceive enough value to pay.

  • Need for audience fit + marketing iteration: he concludes demand is likely audience-dependent, so he’ll focus on marketing next episode rather than expanding product automation.

  • Product implication: offer structure may need to shift toward what users believe is “worth it,” such as:

    • fewer outputs
    • a “confidence” angle
    • tighter expectation management

Frameworks / playbooks explicitly or implicitly used

  • MVP validation / “sell before building”
  • Funnel + conversion thinking: ad/entry → trust → pay link/waitlist gating
  • YC-inspired iteration loop: build minimal fake version → sell → use results to decide next steps
  • Open-model fine-tuning approach: LoRA for likeness + seed variability management
  • Human-in-the-loop value proposition: customers pay for curated best result, not just generation

Presenters / sources

  • Presenter: Joma (creator; “Joma” named in the video intro)
  • Referenced external source: Y Combinator (YC) advice (mentioned, not shown as a presenter)
  • Referenced tooling/source material: Hostinger (sponsorship mention)
  • Model/tool reference: Austris AI (as a source for his evaluation process)

Original video