Real-time inference means running a trained model (a model that has already learned from your data) to produce a prediction in the exact moment a shopper acts, usually in tens of milliseconds, and using that prediction to change what they see right then. A click, an item added to a cart, a payment tapped: the model scores it instantly and the store responds in the same breath. That is the whole idea, and it is what separates a recommendation that adapts as someone browses from a report that tells you what sold last week.
Most Indian retailers already do the slow half of this. Their reports run overnight and tell them what happened. The fast half, acting on a live event before the shopper leaves, is still largely untapped, and it is where the money in modern retail increasingly sits. India’s retail market is on track to reach roughly US $2 trillion by 2032 (Boston Consulting Group and Retailers Association of India, 2024), and the customers driving that growth expect the kind of in-the-moment experience that only real-time decisions can deliver.
Real-time inference vs batch scoring
This is the distinction that matters, so it is worth being precise.
- Batch scoring runs a model on a schedule, often overnight, over a whole pile of records at once, and stores the answers for later. “Which customers are likely to churn this month?” is a perfect batch question. The answer is just as good tomorrow morning as it was at midnight.
- Real-time inference runs the model on a single live event the instant it happens and returns an answer fast enough to act on before the moment passes. “What should I show this shopper who just added rice to their cart?” is a real-time question. By tomorrow, the shopper is gone.
Same trained model, two ways of serving it. The reason real-time is harder is not the maths. It is that the data the model needs (this session’s clicks, this item’s current stock, this card’s recent behaviour) has to be clean, current, and available in milliseconds, every single time. Get that wrong and the prediction is either late or based on stale numbers, which in a live flow is worse than no prediction at all.
The Indian retail decisions that actually need real-time
Not every decision needs to be instant. Real-time earns its cost only when the decision dies if it arrives late. In Indian retail, a handful do.
In-session personalized recommendations. When a shopper is browsing right now, the best moment to suggest a relevant item is right now, scored against what they have looked at this session, not against last month’s segment. Web performance research is blunt about why speed matters here: Amazon found that every 100 milliseconds of added latency cost about 1% in sales (Greg Linden, Amazon, 2006), a finding that has held up across the industry for years.
Dynamic pricing. During a festive surge or a competitor’s flash sale, the right price changes through the day. A model that scores demand, stock, and competitor prices in real time can set the price shown at the moment of viewing, instead of a price decided last night that is already wrong. This is close to demand sensing; see our work on demand forecasting.
Real-time fraud and payment-risk checks. A payment either clears or it does not, and the check has to happen before it clears. This is urgent in India: digital payment fraud jumped more than fivefold to about Rs 14,570 crore (roughly US $175 million) in the year to March 2024 (Reserve Bank of India, 2024). A model that scores each transaction’s risk in milliseconds can flag a suspicious payment in the flow, not in a report read the next morning.
Live inventory and substitution for quick commerce. India’s quick-commerce market reached about US $3.05 billion in FY2024 and is growing more than 40% a year (multiple industry reports, 2024). When delivery is promised in ten minutes, stock changes by the second. If the chosen item just sold out, a real-time model can suggest the substitution a shopper will actually accept, before they abandon the order.
Cart-abandonment intervention. Roughly 70% of online carts are abandoned (Baymard Institute, 2024). A model watching the live session can spot the signals of an imminent drop-off and trigger a nudge, a reassurance on delivery, a small incentive, while the shopper is still on the page, rather than in an email sent hours later.
The common thread: each of these is a decision that depends on your own data and loses its value within seconds. That is exactly the kind of work a custom data and model setup is built for, and exactly where a generic dashboard cannot help, because by the time you read the dashboard, the moment is over.
How a real-time inference engine works, in plain terms
You do not need to be technical to understand the moving parts. There are three, and a live event flows through all of them in well under a second.
Low-latency model serving. The trained model sits behind a fast endpoint, a request comes in, a prediction comes back. The goal is a tight latency budget, often 100 milliseconds or less end to end, because beyond that the experience stops feeling instant. This is the piece our real-time inference engines work delivers.
A feature store. The model needs current numbers to score against, and computing them from scratch on every click is too slow. A feature store keeps those values precomputed and fresh, so the model can grab “this shopper’s last ten views” or “this item’s stock right now” in a moment. It is the difference between a model that knows what is happening now and one guessing from old data.
The speed and cost trade-off. Real-time costs more than batch. You are paying to keep infrastructure ready to answer instantly, all day, instead of running a job once at night. So the honest rule is to make real-time only the decisions that need it. A nightly reorder plan should stay batch. An in-session recommendation has to be real-time. Spending real-time money on a decision that could wait until morning is the most common way these projects waste budget.
Why it is still untapped, and how to start
If the payoff is this clear, why have most Indian retailers not done it? Two real barriers, and neither is the model.
The data is scattered. A typical retailer’s data lives in a billing system, a website or app, a WhatsApp catalogue, supplier emails, and spreadsheets. There is no single, current view to serve a model from, so even a good model has nothing reliable to read in real time. This is the part most “AI” pitches skip, and it is the part that actually decides whether real-time works.
Latency and cost are real constraints. Building something that answers in milliseconds, every time, takes engineering and ongoing spend. On the wrong decision, that is money lost. On the right one, it pays for itself quickly.
The way through is the order of operations, not a bigger budget. First, get the data into shape: collect and clean the scattered sources through data engineering so there is one trustworthy, current view to serve from. This is the audit-first principle in practice, prove the data can support the decision before building the serving layer on top.
Then be selective. Pick one decision that is clearly time-sensitive and high-value, a recommendation, a fraud check, a substitution prompt. Run a small proof on that single decision, measure whether it moves the number that matters, and only then expand to the next. You do not have to make everything real-time. You have to make the right thing real-time, and prove it.
How Galific approaches real-time inference
We start with the data, not the demo. A low-cost data audit tells you whether your sources can support a real-time decision before you spend on serving infrastructure. From there it is one focused decision, served behind a fast endpoint, wired into the channel your shoppers actually use, an app, a website, a quick-commerce flow, and monitored for both speed and accuracy so it keeps earning its keep. It is delivered from India and priced for SMEs, and it sits inside the wider retail and e-commerce work, because a real-time decision is only ever as good as the data underneath it.
The untapped part of Indian retail is not more reports about yesterday. It is the decision made in the moment the shopper is still deciding, one clean data foundation, one model served fast, one action that lands before the moment is gone.
