Case study · Ken AI
How Ken AI eliminated inference bottlenecks
and cut batch runtimes by 68%
68%
faster batch runs (8h to 2h 35m)
39%
reduction in batch failures
~half a day
integration with no meaningful code change
Background
Ken AI runs high-volume, individually-written cold email for B2B SaaS companies - personalizing, rewriting, qualifying, and segmenting copy so each prospect gets a genuinely bespoke message rather than a template. That happens at serious scale: 1M+ emails sent per month across 20+ B2B SaaS clients.
Problem
Ken AI's personalization engine is only as good as the inference behind it - and the hosted inference APIs they previously relied on couldn't keep up at scale.
Large jobs would stall or fail partway through, so the team was left stitching together multiple third-party providers and routing each batch through whichever one happened to be available. When a run died mid-job, campaigns slipped: days of launch delays, engineers pulled in to babysit jobs, and frustrated clients waiting on sends.
“Large jobs would stall or fail partway through, and we couldn't trust them to finish a run on time - every slipped send turned into a frustrated client.”
- Cristian Frunze, Founder, Ken AI
Solution
Because Doubleword is OpenAI-compatible, Ken AI pointed their existing batch pipeline at it and had it running in about half a day with no meaningful code changes. The standout was Doubleword's async tier, which let them fire off huge jobs and collect results without holding connections open, paired with a UI that made runs genuinely easy to watch and manage.
Just as importantly, Doubleword gave the team reliable access to leading open-weight models without needing to manage infrastructure or juggle multiple providers. The payoff was immediate: average batch run time dropped from roughly 8 hours to 2 hours 35 minutes - and, more importantly, jobs completed far more reliably.
“Doubleword is the first inference provider that just runs our batch jobs reliably and fast, with a clean async API and a UI that actually helps. It took half a day to integrate and immediately cut our batch times by more than half.”
- Cristian Frunze, Founder, Ken AI
Results
-
Batch runs 68% faster - average run time fell from ~8 hours to 2h 35m.
-
39% reduction in batch failures - runs complete reliably, so campaign-launch delays are largely gone.
-
Engineers off babysitting duty - the team no longer monitors inference jobs manually, campaigns ship faster, and they have dependable on-demand access to open-model inference for testing and swapping models.
-
One platform for large-scale inference - supporting the generation of more than 1 million personalised emails per month.
What's next
Ken AI is steadily moving more of its workloads onto Doubleword - personalization, qualification, segmentation, and reply handling - with the goal of running the bulk of its inference on the platform in the near future.