Government · ML · OSINT
Spotting coordinated hostile actors in live social media — at a volume no analyst team can read.
I co-founded Patagona Technologies to build machine-learning and open-source-intelligence platforms for public-sector intelligence work: real-time ingestion, graph-based detection of coordinated activity, and the dashboards analysts use to act on it.
- Client
- Patagona Technologies (co-founded)
- Role
- Co-founder & CTO
- Period
- 2018 – present
- Focus
- Government intelligence
- co-founder and CTO
- 8 yrsco-founder and CTO
- social media stream analysis
- Real-timesocial media stream analysis
- model retraining on Kubernetes
- Continuousmodel retraining on Kubernetes
The problem
Hostile campaigns move through social media faster than people can read it, and their tactics change constantly. Models trained once go stale within weeks.
Analysts only act on output they trust — so detection had to arrive in tools that show their working, not as a black-box score.
What I built
- 01
DEPRAM — hostile-actor detection
Led a team of engineers and data scientists building graph-embedding classifiers, and the infrastructure to run them against real-time social media streams.
- 02
Continuous MLOps
Kubernetes pipelines on Apache Beam, Argo Workflows and Airflow covering ingestion, feature engineering, retraining and inference, so models keep pace with the data.
- 03
THREATDESK — threat-scoring dashboards
Node.js microservices behind an API gateway on Kubernetes, fed by OpenSearch ingestion pipelines, giving analysts real-time threat scores they can drill into.
- 04
Knowledge and OSINT platforms
Search and retrieval across large document collections, with retrieval-augmented workflows that surface the passages that matter.
The result
- Eight years leading engineering for public-sector intelligence work, from prototype to production.
- The disciplines it demands — continuous retraining, auditability, real-time pipelines — are the ones I bring to every AI system I build.