Most AI teams optimize a metric. This one has to build something a city council, an oversight body and a court can all look at afterwards, which is a harder and more interesting constraint.
The Company
IREX is a distributed team of roughly 100 people, more than 50 of them engineers and data scientists, with executive leadership in the United States, Andorra and the UAE and regional delivery teams across Latin America and Europe. Engineering covers back end, front end and UX, DevOps and network operations, QA and support.
The work is in production in eight countries on more than 250,000 cameras, which means the systems you build are used by police officers, transit operators and airport security on live incidents. That is a strong reason to care about false positives.
Everything runs on an open-source stack, on customers’ own hardware and IREX-authored components are available for source review or escrow. If you dislike building things nobody outside the company can inspect, that will suit you.
Open Roles
We hire slowly and we publish a role only when it is genuinely open. This is what is open now.
Where We Hire
Back end, front end and UX, DevOps and network operations, QA and support, on Kubernetes and an entirely open-source data plane.
Computer vision, Vision-Language Models, dataset construction including synthetic environments and bias measurement.
Solution architecture, site surveys, capacity modeling and delivery on national-scale programs.
Public-sector business development, pre-sales engineering, bid and proposal work and regional partnerships.
Governance, information security and the audit and disclosure machinery the platform is built around.
Media relations, social and video production, across a global public-sector audience.
What We Hold To
Keep Reading
We are more interested in what you think is wrong with public-safety AI than in a list of frameworks.
Two things arrived at once. Language models made a detector cheap to build, and regulation turned accountability into a purchasing requirement: the EU AI Act, the NIST AI Risk Management Framework and the FBI CJIS Security Policy all ask an agency to show how a decision was reached. A platform already built around Case ID, an audit trail and customer-owned data starts that conversation somewhere different from one retrofitting them.
Both, and the second follows from the first. AI agents work across engineering, security, documentation, marketing and sales support, grounded in a version-controlled knowledge base that agents themselves maintain. Each one runs under a named human owner, tool allow-lists and human approval before anything irreversible, governed by a board-approved policy. You would be expected to work that way, not to watch it happen.
Nothing in the code, and we do not pretend otherwise. As a growing share of software is machine-generated, features and algorithms replicate in a fraction of the time they used to. What does not replicate is a deployment record across eight countries, the trust of a government buyer and an architecture an oversight body can audit. That is why the constraints are engineered into the platform rather than written into a policy.
Because here they are the same decision. A deployment that cannot show a lawful basis for a search does not survive a council meeting, a court or a procurement review, and an agency asked to cover more with the staff it already has cannot afford a system it has to defend twice. Narrow constraints, the Case ID mandate and the audit trail are engineering work, and they are why deployments stay in production.
The team is distributed by default, with leadership in the United States, Andorra and the UAE and delivery teams across Latin America and Europe. Specific roles state their location and travel expectations.
Entirely open source: Kubernetes, Ceph, Cassandra, Kafka, PostgreSQL, Elasticsearch and Apache Spark, with a proprietary CPU inference framework and an NVIDIA CUDA GPU pipeline for analytics. No third-party proprietary components are required.
