How MangalyaMatrimony's AI matchmaking works

The signals we use

Every signal below is a structured field on the profile — explicitly entered by the user or inferred from interaction. No free-text scraping, no inferred caste from name, no demographic guessing.

How AI and Guna Milan combine

We treat AI ranking and traditional horoscope matching as two independent signals shown side by side on every match — not a single blended number. A family can choose which one to weight more, or use both.

1. Hard filters first: Caste, sub-caste, gotra exclusion, mother tongue and any partner-preference deal-breakers narrow the pool. The AI never overrides these.
2. Horoscope filter (optional): If you set a minimum Guna Milan threshold (default 18, configurable to 24+ or 32+), profiles below the threshold are hidden. More on Guna Milan thresholds.
3. AI ranks the survivors: Within the filtered pool, the 44-signal model ranks profiles by predicted compatibility. The score is shown as a 0–100 percentage on every match card.
4. Both scores visible: The match card shows the AI compatibility score and the Guna Milan score (out of 36) so families can decide which to act on.

What we never use

• No race / skin-tone scoring. Complexion is captured as a legacy preference field for backward compatibility but is never used to rank or filter matches in the AI model.
• No chat content or message text is fed into the matchmaking model. Engagement signals are frequency and recency only, never the content of conversations.
• No photos are used for AI scoring beyond the safety / authenticity check. We do not infer attractiveness, ethnicity or expression from photos for matchmaking.
• No off-platform data. We do not buy demographic data, do not import social-graph data, and do not enrich profiles with third-party data brokers.
• No model training on identifiable data. Only anonymised aggregate signals (accept / decline rates per cohort) update the ranking model. Personal identifiers, photos, contact details and chat messages are never used for training.

How we evaluate the model

We measure the model on three outcomes, in order of importance:

1. Interest-acceptance rate — the share of AI-suggested matches who accept a sent Interest. A higher rate means the suggestions feel right to both sides, not just the sender.
2. Time-to-shortlist — how long an active user takes to build a working shortlist of 5–10 mutual-Interest matches. Lower is better.
3. Match-to-wedding conversion — the share of mutual-Interest matches that proceed to family meetings and eventually to marriage. Reported quarterly to the team; surfaced annually in our public stats.

The model is retrained at most monthly, only on anonymised cohort-level signals, with a held-out test set covering all four South Indian languages and the major NRI segments. A version of the model never goes live without passing offline A/B against the previous version on all three metrics for every major sub-community.