Dira
AI real estate intelligence
Location intelligence for property decisions — scoring, land cautions and commute insight over a spatial data engine.
- NestJS
- Next.js
- PostGIS · H3
- Python geo-ml
- MapLibre
- LLM
Context
Dira is for people making property decisions in Kenya — buyers, renters, and investors — who need to understand a *location*, not just a listing. Today that judgement comes from an agent's word, a drive-by, or scattered, unreliable data. (In development — this page describes what I'm building toward.)
The problem
Kenyan property data is fragmented and inconsistent, and the hard part of a property decision isn't the building — it's the location: commute reality, nearby amenities, and land-related cautions. Listing portals index properties; they don't reason about the ground underneath them.
Approach
The architecture I'm building toward: a spatial data engine that aggregates signals per area using PostGIS with H3 hexagon indexing, so every location is scored on a consistent grid rather than by fuzzy neighbourhood names. A Python geo-ml service is planned to produce the location scores, with an LLM layer to turn that spatial data into plain-language insight — scoring, land cautions, and commute context. NestJS for the API; Next.js + MapLibre for the map. The decisions I'm committing to early: H3 for uniform spatial aggregation, and keeping the geo-ml service separate from the API so models can iterate independently.
Securing it
Location intelligence is a trust product — a wrong "land caution" has real consequences — so I'm designing provenance and data integrity in from the start, not bolting them on. The plan: validate source data, scope and rate-limit the LLM layer so it can't be steered off its data, and qualify land-caution outputs rather than assert unverified claims as fact.
What shipped
In development — currently repo scaffolding. Nothing to demo yet; this page is the design I'm building toward.
Links