Work · SafeTrekr · 2025–2026 · Production
TarvaRI
Public risk feeds turned into trip-specific alerts.
Availability
TarvaRI runs in production as infrastructure behind SafeTrekr, so it has no public link of its own.
The problem
Hazard feeds arrive in different formats, geographies and severities, most items have nothing to do with a given trip, and an alert that arrives late is only noise.
The system
A production intelligence pipeline: source connectors and custom parsers, a normalized store, clustering by place and time, a risk engine, LLM-assisted triage with human review, and routing into SafeTrekr trips.
Outcome
TarvaRI turns public hazard feeds into alerts tied to the specific SafeTrekr trips they affect, so analysts review what matters instead of watching every feed.
Proof
It runs in production inside SafeTrekr, with evaluation, replay, shadow mode, worker monitoring and drift detection built in.
Role
Tarva conceived TarvaRI and architected and led its AI-assisted development, including information architecture, UX/UI, brand and design. EIGHT32, our engineering partner firm, built the final bridge into SafeTrekr.
Supporting stack
Python · FastAPI · Supabase · pgVector · Redis · Ollama · React console
Scale
40 active source connectors · 17 categories · 43 custom parsers
