AI real-estate deal finder case studyAn AI deal-finder that ran at roughly 75% lower cost than the original approach.
A founder wanted to surface promising real-estate deals from a constant stream of listings — and the early version was expensive to run.
The real product, shown with demo data.
A founder wanted to surface promising investment deals automatically, but the early version was expensive to run. I rebuilt it as isolated pipeline stages with a rules-first, AI-second cascade — cutting running costs by roughly 75% while keeping quality high.
Promising deals, buried in a stream of listings.
A founder wanted to automatically surface promising real-estate investment opportunities from a constant stream of new listings — and the early version was expensive to run.
Scale the volume without the AI bill spiralling.
Two things at once: collect large volumes of listing data reliably, and use AI to judge which deals were worth attention — without the AI costs spiralling as volume grew.
A pipeline of isolated, fault-tolerant stages.
A data pipeline built as isolated stages — collection, AI enrichment/scoring, and delivery — running as independent background workers off a queue, so one stage failing never took the rest down. Screens included a listing analyzer, deal scores, a scenario builder and a subscriber dashboard.
Screens from the real product, captured with demo data. No client or customer data is shown.
Isolated stages, running off a queue
- 01CollectListings from a constant stream
- 02Rules firstEverything knowable from the data
- 03AIAI secondOnly the judgment calls
- 04DeliverScored deals, alerts, dashboard
Rules first, AI second.
Rules-first cascade
Deterministic code handled everything knowable from the data; AI was reserved only for the judgment calls.
Dedup + caching
Together with the cascade, cut the running cost by roughly 75% while keeping quality high.
Fault isolation
Failures were contained and retried cleanly, stage by stage.
The calls that shaped the build
Deterministic code is cheap, fast and predictable. Reserving AI for genuine judgment calls — and adding deduplication and caching — is what cut running cost by roughly 75%.
So one failing stage never takes the rest down. Failures are contained and retried cleanly.
The client later wound the product down for business reasons unrelated to the build. It is shown as engineering work delivered, not an active product.
Quality deals, surfaced far more cheaply.
Sikandar was an absolute pleasure to work with. He is professional, experienced, and brought valuable insights to the development of my platform. He wasn't afraid to offer constructive feedback and push back when it would improve the project, including steering me away from spending more than necessary. After reviewing the data and pricing, it's clear he understood the challenges long before I did and genuinely cared about my budget. His expertise and collaborative approach made the process smooth and efficient. Highly recommended!
