Research Engineer

Atul Saxena

I build the environments models are trained in, and the tools that show what they learned.

RL Systems
Interpretability
ML Infrastructure

Projects— projects/

Projects

Security2026

Meridian

Agentic OSINT geolocation on 4 GB of VRAM

Source

An agentic geolocation engine with eight signal extractors and Bayesian elimination, producing fully auditable investigation reports. Runs locally, no API keys.

//The problem

Geolocation tooling in this space is either a cloud service you hand your data to, or a manual process with no audit trail for how a conclusion was reached.

//What it does

Eight independent signal extractors feeding a Bayesian elimination process.
Every conclusion traced back to the signals that produced it.
Runs on 4 GB VRAM with no external API dependencies.

//Measured

Footprint: 4 GB VRAM, no API keys
Extractors: 8 signal types

//Built with

Python
Bayesian inference
Agents
OSINT

status: Active