MyPhotoAI

AI headshots from your selfies.

An AI headshot generator that turns selfies into professional photos for LinkedIn, resumes, and dating profiles. A free first photo and a preview Lookbook drive paid unlocks through Stripe, fed by a programmatic SEO engine that brings the traffic.

In plain EnglishUpload a few selfies, get back a gallery of professional-looking photos of yourself. You see free watermarked previews first, so you know the photos actually look like you before paying to download the clean versions.

Lead engineer and operator.

MyPhotoAI running in production — live at myphotoai.io

The exhibits

Drag to spin. Click any screenshot to inspect it full size. Every one is the live product.

MyPhotoAI landing page — The landing page: a drag-to-spin 3D ring of looks. The product demos itself before asking for anything.Drag to spin · click to inspect · 1/5

The problem on file

AI headshot products die two deaths: users who never believe the photos will look like them, and unit economics where compute is spent on users who never pay. MyPhotoAI's answer is a proof-first funnel: generate free watermarked previews so the user gets the that's-me moment, then convert that into a paid HD unlock. Which means the entire business rests on one technicality: the free previews must be genuinely free to see and genuinely impossible to steal. Most products fake this with a CSS blur or a canvas overlay, and DevTools defeats both in a minute. Here the paywall is the storage architecture itself.

What it is, in numbers

programmatic SEO pages live

1,898

photo styles, CI-enforced registry

127

generation providers, lane-routed

local currencies at checkout

free browser tools as SEO surface

lines of TypeScript across 4 services

84k

The marquee piece

Clean images and previews never share a bucket. The generation worker uploads the clean image to a private bucket, a separate watermark service composites the preview into a public bucket, and the database row the client sees only ever contains the preview URL. The clean path lives server-side.

Unlocking is then a payment fact, checked where it can be enforced: a 402 at the only place clean bytes can leave the system. There is nothing to bypass in the client because the client never holds the clean URL.

The free previews themselves are ordered by face-similarity score, so the four clear shots a skeptical visitor sees first are the strongest that's-me moments the batch produced. The conversion engine and the security model are the same design.

Architecture

The paywall is the storage topology: clean bytes live where the client cannot reach.

The hidden ninety percent

The hidden ninety percent· what most portfolios skip

01Generation is lane-routed across four providers: Gemini model variants for most categories, InstantID on free Kaggle T4 GPUs for categories Gemini's safety filter rejects, and Flux LoRA on Replicate or self-hosted RunPod. The free lane doubles as a zero-dollar failover when the spend gate trips.
02The spend gate learns. Daily Gemini budget is a Redis counter whose soft cap is the trailing 7-day median of delivered images; median, not mean, so an abuse spike cannot teach the baseline to tolerate an attack. It fails open on transient Redis errors and closed on Redis OOM, and pages Telegram either way.
03Safety blocks get a 3-tier recovery ladder: substitute known trigger phrases, retry, then fall back to a fully neutral prompt. Research in-code concluded no wording reliably passes the output filter; only removing the trigger does. Input-side blocks skip retries entirely, because their recovery rate was zero.
04Prompts are model-aware. Models that over-smooth skin get an anti-plastic clause; models that render skin natively get a leaner prompt because over-stuffing degrades them. A skin-tone-fidelity clause exists because a regression was observed, dated, and fixed.
05The pSEO engine gates spend the same way the product does: cheap LLM content passes a helpfulness gate before expensive image generation runs, layouts vary deterministically per slug so no two pages share a skeleton, and the IndexNow queue liveness-checks every URL before submitting it.
06High-fraud regions get a pay-first gate in 10 local currencies: a small payment before the first generation, redeemable for photos after. Compute never runs for a user who has proven nothing.
07One SPA-mode misconfiguration once hid four months of broken tool-page assets behind soft 200s. The fix is a build-blocking guard: every route must be prerendered or explicitly redirected, or the build fails.

In production

01Revenue is real and third-party verified: Stripe-audited figures are public on TrustMRR (trustmrr.com/startup/my-photo-ai-inc).
021,898 static pages serve from Cloudflare Pages at near-zero cost, built from a 2,158-keyword manifest with an E-E-A-T publish bar.
031,094 commits since taking over the codebase in February 2026, with Grafana and Loki monitoring and two Telegram bots separating ops alerts from exec revenue digests.

Take away

The conversion funnel, the cost controls, and the paywall are one system, and it is all server-side.

The stack

ReactViteTypeScript Supabase Google Gemini StripeIMG.LY background removalMediaPipeDockernginx-proxy

Skills it proves

PaymentsStripe integrationStripe Checkout and webhooks driving paid conversion in a consumer product.

AIGemini image pipelineA production image flow on Gemini with on-device background removal and vision.

GrowthProgrammatic SEO at scaleData-driven page generation with sitemaps, IndexNow, and intent funnels.

GrowthIntent-driven conversion funnelsMatch search intent to a paid action.

InfraDocker and self-hosted deployContainerized services behind a reverse proxy with automatic TLS.

BackendSupabaseAuth, Postgres, and row-level security for a consumer app.

Next step

Want this for your product?

Send a short brief. I read every one and reply myself.

Send a brief [email protected]