How AI Matchmaking Actually Works — And What It Cannot Do

The phrase "AI matchmaking" is doing a lot of unexamined work in Indian matrimony advertising right now. It suggests a machine that understands love and identifies the person you are meant to marry. Nothing of the sort exists, and any platform implying otherwise is selling you a story.

What does exist is genuinely useful, just less mystical. A matching system is a ranking engine. Its job is to take a pool of tens of thousands of profiles and decide which twenty you should look at today. It cannot tell you who to marry. It can substantially change the odds that the twenty profiles in front of you are worth your evening.

The problem it is actually solving

Traditional matrimony search is a filter query. You set age 26 to 31, city Bengaluru, community, education, income band — and the site returns everyone who satisfies the conditions, usually sorted by whoever logged in most recently. Two things go wrong with this.

First, filters are binary and human preferences are not. Someone a year outside your age band or in the next city is excluded entirely, even if they are otherwise an unusually good fit. Second, recency ordering is not relevance ordering — it surfaces the most active users, not the most suitable ones, which is why the same handful of profiles keep reappearing.

A ranking system replaces "does this profile qualify?" with "how well does this profile fit, relative to every other option?" That single shift is most of what people mean by AI matchmaking.

What the system actually looks at

Broadly, four categories of signal, in increasing order of usefulness:

1. Stated attributes

Age, location, education, profession, income band, diet, community, family type, languages. This is the same data a filter uses. The difference is that it is scored on a gradient rather than applied as a gate — a near miss on one attribute reduces the score instead of deleting the profile.

2. Stated preferences

What you said you are looking for, and — importantly — what they said they are looking for. Mutual fit matters far more than one-directional fit. A profile that matches your criteria perfectly but for whom you fall outside their criteria is close to worthless to show you, and a great deal of wasted effort on matrimony sites comes from exactly this asymmetry.

3. Astrological compatibility, where you want it

Where both profiles carry birth details, Ashtakoota (36 guna) compatibility and Mangal dosha status can be computed and used as an input. This is deterministic astronomy plus classical rules, not machine learning — the Moon's nakshatra at a given time and place is a calculation, not a prediction. It is included because a very large number of Indian families will ask for it eventually, and finding out at the shortlisting stage is considerably less painful than finding out after three months of conversation.

4. Revealed behaviour

This is the signal with the most information in it, and the one people underestimate. What you actually open, send interest to, reply to and continue talking to is a far better description of your preferences than what you typed into a preferences form. Almost everyone states one thing and behaves differently — people write "height 5'8'' and above" and then have their longest conversations with someone who does not meet it.

Why a mutual score beats a one-sided one

A matching system that only optimises for what you want produces a queue of people who will never reply. The useful quantity is the likelihood that a conversation happens and continues — which requires scoring both directions and weighting them together. This is also why a good system will sometimes show you someone slightly outside your filters: it has reason to believe the interest will be returned, which is worth more than a perfect attribute match that goes nowhere.

What AI matchmaking cannot do

This part matters more than the capabilities, and it is the part the advertising leaves out.

• It cannot predict whether a marriage will last. Nothing observable at the profile stage does. Marriages succeed or fail on conflict handling, financial stress, health, family dynamics and events years in the future — none of which are in the data.
• It cannot assess chemistry. Whether two people enjoy each other's company is only discoverable by those two people, in conversation. No score substitutes for that.
• It cannot detect character. Honesty, kindness and how someone behaves under pressure are not fields in a database. A verification badge proves identity, not decency.
• It cannot fix a dishonest profile. If the inputs are inflated — height, income, marital history — the ranking is confidently wrong. This is why identity verification matters as much as the algorithm.
• It cannot know what you have not told it. A half-filled profile with two lines of text and one photo gives the system almost nothing to work with, and you will get correspondingly generic results.

The fairness problem, stated plainly

A system that learns from behaviour will learn whatever biases are present in that behaviour. In the Indian context that includes caste, colourism and income prejudice, and a naive engine will happily amplify all three because they are statistically predictive of who replies to whom. Treating that as a neutral technical outcome is a choice, not an inevitability.

The countermeasures are unglamorous and mostly consist of restraint: not inferring sensitive attributes that a user has not chosen to state, not letting behavioural signals silently narrow a pool the user never asked to narrow, keeping community and religion as preferences a person sets deliberately rather than things the system decides for them, and keeping a visible diversity of profiles in every result set rather than converging on a single stereotype. Any platform claiming AI matchmaking should be able to say what it does here. Most cannot.

How to get better results from any matching system

• Complete the profile fully, including the unglamorous structured fields. Every blank is a signal the system does not have.
• Write specific free text rather than adjectives. "I teach Class 9 physics and coach the school debate team" carries information; "I am family-oriented and fun-loving" carries none.
• Add accurate birth details if Kundli matching matters to you or your family — an approximate birth time can move the Moon into the wrong nakshatra and change the result.
• Keep your stated preferences honest rather than aspirational. Filters set to what you wish you wanted shrink your pool and teach the system the wrong thing.
• Actually use the platform. Opening, shortlisting and replying are how the ranking improves. A dormant account gets a static, generic feed.

The realistic summary

AI matchmaking is not a prediction of your future. It is a better ordering of a very large list, built from stated attributes, mutual preferences, optional astrological compatibility and the difference between what you claim and what you do. That is a meaningful improvement over a filter query sorted by last login, and it is worth having. It is not a substitute for talking to someone for three months, asking uncomfortable questions and forming your own judgement.

On Sangamly the compatibility score is shown as information rather than as a verdict, alongside verification status and Kundli details where both people have provided birth data. The system narrows the field. You decide. That division of labour is the only one that actually works.