Intterra

Real-time geospatial alerting for fire, EMS & law enforcement. Two-tier architecture — Redis GEO hot layer over PostGIS — with Kafka burst absorption and LLM incident summaries gated by LangSmith evaluation.

The Challenge

Fire, EMS, and law-enforcement agencies needed real-time situational awareness across 100k+ GPS events a day — with proximity alerts fast enough to matter mid-incident. A single database layer couldn't hold latency under burst load. Standard request-response architecture couldn't meet the bar: proximity alerts only matter if they arrive while the incident is unfolding, and dispatch data comes in unpredictable bursts — a quiet Tuesday and a wildfire evening differ by orders of magnitude.

Our Approach

A two-tier architecture: a Redis GEO hot layer over a PostGIS system of record, with Kafka absorbing burst loads. LLM incident summaries are gated by LangSmith evaluation before anything reaches a responder's screen. Around that hot path we built the supporting system: Airflow ETL pipelines unify multiple emergency data feeds into one operating picture, and structured prompt workflows generate incident summaries that must pass LangSmith evaluation before any responder sees them.

What We Built

A proximity-alert engine — Redis GEO hot layer over a PostGIS system of record; Kafka ingestion absorbing burst loads across 100k+ GPS events per day; LLM incident summarization gated by LangSmith evaluation before responders see it; Airflow ETL pipelines unifying multiple emergency data sources into one view; and structured prompt workflows for contextual incident insights.

Key Decisions

Two-tier geo store — proximity matching runs against a Redis GEO hot layer while PostGIS stays the system of record: sub-500ms reads under load without sacrificing durable geospatial queries. Eval-gated LLM output — incident summaries pass LangSmith evaluation before display; responders making life-safety decisions never see an ungated model output. Burst absorption by design — major incidents create traffic spikes by definition, so Kafka decouples ingestion from processing and the platform degrades gracefully instead of dropping events.

The Outcome

Sub-500ms proximity matching at 100k+ GPS events a day, serving 10k+ concurrent responders — with an embedded Nugen team owning the platform since 2024. The engagement continues as an embedded dedicated team — the same engineers who designed the architecture run it, extend it, and carry the pager.