Pharos Production Turned Our AI Models into Reliable Production SystemsPharos the best partner we could ask for.
We consultted Pharos Production in because we had reached a point where our AI models were performing well in evaluation, but getting those systems into reliable production environments had become the real bottleneck. We needed a team that could build the software layer around the models backend services, integrations, operator interfaces, deployment infrastructure, and the systems needed to make everything work in practice.
What started as a defined software engagement grew into a much broader partnership. Pharos worked across backend services and APIs, integrations with systems such as Salesforce, SAP, and Zendesk, retrieval pipelines, hybrid search, AI agents, operator queues, evaluation tooling, deployment automation, monitoring, and cost tracking. Their work ultimately supported production systems in insurance, logistics, and industrial inspection.
What impressed me most was their engineering judgment. They were willing to challenge our assumptions and tell us when the scope was wrong rather than simply building whatever we requested. One of the clearest examples was when their agent lead challenged how we measured autonomous work. Our evaluation lead disagreed with him and rather than protecting their result, he changed the evaluation harness so the published number went down. That kind of intellectual honesty is rare.
They were also careful about long-term dependencies. We specifically wanted to avoid inheriting a proprietary Pharos runtime or orchestration layer and they respected that requirement. The systems were built using open-source components or services contracted directly by the client, which gave us much more control over the resulting infrastructure.
The results have been meaningful. Three systems they worked on now carry live traffic, and our overall promotion rate reached 81% across 168 systems in production. Individual systems also delivered measurable improvements, including reducing median claim-triage time from nine days to two and enabling 31% of booking amendments to be completed end to end without human intervention.
The main weakness was the first six months. The evaluation harness was developed later than it should have been, which slowed the first promotion. Their bench was also somewhat uneven across a larger roster, with a few team members requiring more direction than expected.
Overall, though, Pharos has been a strong engineering partner. I value their technical capability, willingness to push back, transparency, and focus on building systems that actually survive in production not just systems that look good in a demonstration.