Amorelle.

The problem

The dating-app era is ending. After a decade of growth, the swipe-to-match market has gone ex-growth — US dating-app downloads fell ~14% and monthly actives ~8% year-over-year in late 2024[5], and the products that are growing are the relationship-oriented ones (Hinge grew ~23% in a quarter while Tinder declined)[6]. The center of gravity is shifting from helping people meet to helping couples stay connected.

But the tools couples actually use are thin. The category leader, Paired, is expert-content-driven — not AI-native, not personalized, not localized — and users routinely complain its prompts feel generic. Amorelle was built for the couple the incumbents underserve: one that wants something that feels made for them, in their language.

Market & opportunity

There is no clean tier-1 figure for the "couples app" sub-segment — so the honest anchor is the broader online dating application market: ~$7.9B in 2022 growing to ~$14.4B by 2030 (7.6% CAGR)[1], with the demand inside it visibly migrating from acquisition to retention. That migration is where the capital is going: Arya raised $21M in March 2026 for an AI "relationship OS"[4], and the adjacent AI-companion category is on track for ~$120M in 2025 revenue with downloads up 88% year-over-year[7] — proof that emotionally intelligent AI is sticky.

Two structural levers are under-exploited by incumbents. First, personalization — the strongest retention lever in a category where the leader ships one-size-fits-all content. Second, localization: roughly 72% of apps are English-only, yet localized apps see materially higher downloads and revenue per market[8]. Amorelle attacks both — AI-native personalization and eight locales including RTL — aimed first at the non-Western couples the category ignores.

Who it's for

Couples in an established relationship (the research segments them as sticking vs stalled), cohabiting or long-distance, who have already adopted at least one shared digital tool together. The first-order risk for any two-person app is partner activation — if only one of the pair joins, the account is dead — so the wedge is as much about the pairing ritual as the feature set.

Constraints

• A two-person app fails if either partner churns. Already-brutal D30 retention compounds: the loop has to give both people a reason to return.
• AI can't be a free LLM proxy. Open-ended client-supplied prompts would be an abuse magnet; all prompts are built server-side from a feature type, and client-supplied prompts are rejected.
• Relationship distress is a safety surface. Mediation and counseling features can surface self-harm or abuse language — those paths need stricter handling than a date-idea generator.
• Multilingual + RTL from commit one. Retrofitting RTL and eight locales into a feature-dense app later is the multi-day tax this build refused.

Architecture & what I built

The two-model AI router

Every AI feature — counselor, mediator, letter and vow helpers, daily spark, advice, quiz generation, date planning, translation — flows through a single ai-router edge function that selects the model by risk, not by feature alone. Routine prompts go to Groq (Llama 3.3 70B) for speed and cost. A red-flag crisis detector runs on the latest user message; when it fires (and a Claude key is present), the request escalates to Claude with a stricter crisis prompt. Cost is bounded server-side: max 24 messages, ~12KB of context, a 6KB system-prompt ceiling, and a per-user daily quota.

Safety as server-side policy

All system prompts are assembled on the server from the feature parameter — the client can't inject its own — and every prompt carries a safety footer (never diagnose, never prescribe, never impersonate a therapist, never take sides, never help one partner deceive the other). The app also runs the crisis regex client-side, so an SOS banner can appear before the server even responds. It's the same safety posture as Coco, applied to a two-person context.

Realtime shared state + the feature surface

Supabase Realtime powers the genuinely shared objects — chat with voice notes, shared albums and "lockets," paired mood check-ins, AI-written memory books and weekly stories — scoped by couple_idwith RLS on private media buckets. Around that core loop sits a deliberately broad surface (cycle tracking, shared calendar, streaks, wishlist, trips, love-languages, games) so that whatever a given couple's "our thing" turns out to be, there's a hook for it. Which hook actually drives shared return is exactly what the research below is designed to find.

Embedded research — the Core-Loop Discovery Study

Rather than guess which feature to deepen next, I designed a qualitative study (with the user-research-skill, constructivist / semi-structured method) to answer the one question that decides a couples app's fate. It is ready to recruit from the Supabase user base.

The study deliberately treats a misunderstood feature as a finding, not a user error — and maps every theme back to those four build decisions. It's the product-discovery counterpart to the engineering: ship a broad, instrumented surface, then let real dyadic behaviour decide where the next cycle goes.

Trade-offs

• Breadth-then-narrow over a single hero feature. Shipping many shared surfaces risks diffuse focus, but it lets the research identify the real core loop from behaviour instead of a founder's guess. The narrowing is a deliberate later step.
• Dual-model routing over one provider.Two providers means more operational surface (keys, fallbacks, prompt parity), but it buys Groq's latency/cost for the 95% case and Claude's judgment for the 5% that matters most.
• Day-one i18n/RTL over English-first speed.It slowed the first build, but it's the lever that opens the structurally underserved markets the deck is aimed at — and it can't be bolted on cheaply later.

Goals & what's next

Outcome

Amorelle is live on amorelle.ai, the App Store and Play Store, with the dual-model router, realtime shared state, server-side safety, and eight-locale support all shipping — and a rigorous research instrument ready to point the next cycle. The thesis: as the market pivots from dating-acquisition to relationship-retention, the winner is AI-native, safety-routed, and localized— not another content library with a couple's name on the cover.