Company
Unstract is an AI-powered enterprise software company founded in 2022, headquartered in California, with a fully remote team operating across India. Its flagship automation product is a no-code, open-source Agentic Document Extraction platform that enables enterprises to automate complex document-based workflows – replacing manual processes that could take days with automated pipelines that complete in minutes.
Kubernetes underpins every customer-facing workload. Unstract runs a cloud-hosted managed service as well as on-premises deployments within customer infrastructure. Given this dual operating model and the compute-intensive nature of LLM inference workloads, reliable, efficiently-sized Kubernetes clusters are not optional – they are, in fact, the product.
The Challenge: From Guesswork and Midnight Pages to Autonomous Infrastructure
As a lean, fast-moving startup, Unstract made a conscious early trade-off: overprovisioning infrastructure to get to market quickly and optimizing later. The strategy worked for growth, but carried compounding technical debt.
Three problems became impossible to ignore as the customer base expanded:
- Node sizing was pure guesswork – Selecting VM types for node groups required rough calculations and intuition rather than data, and teams were often constrained to a single node type per group, further limiting flexibility.
- Workload resource settings lagged reality – Load testing informed initial limits, but production traffic patterns rarely matched those in the lab. New components were often sized by analogy with existing ones rather than by measured need.
- Memory spikes caused real downtime – Stateful components like Redis or MinIO had their request and limit values set identically and pegged to a worst-case high-water mark. When anomalous traffic arrived (including one incident involving automated abuse of a loophole), OOMKill events triggered midnight pages and manual intervention.
We used to get pulled in every day, basically – every time an issue happened. The mileage of folks understanding Kubernetes varies, and sometimes, if it were folks who are not really well-versed with Kubernetes, then definitely I or someone else would get pulled in.
Ritwik G, Staff Engineer at Unstract
Unstract’s engineering team tried to address these issues manually over a three-to-four-month period – load testing, node rightsizing, workload autosizing – but the effort wasn’t paying off. Workloads evolved, new components arrived, and the optimization work had to start over.
Cost optimization used to be an activity that would bring down the cost once, but then it spirals out of control after some time, and then again we had to repeat. It was a one-time activity rather than a continuous one.
Ritwik G, Staff Engineer at Unstract
The team also evaluated stitching together open-source tools – a VPA implementation, spot-node scheduling, and node rightsizing. However, the operational overhead of learning, configuring, and maintaining a bespoke toolchain was untenable for a team without dedicated DevOps headcount.
The Solution: One Platform to Replace a Patchwork of Tools – and a Backlog of DevOps Tickets
Ritwik G, who manages infrastructure at Unstract, discovered Cast AI while researching Kubernetes optimization tools. A closer look confirmed it covered every capability on the team’s roadmap: VPA, automatic node rightsizing, and partial Spot-node utilization – all in a single, Kubernetes-native platform.
The main differentiator was the Kubernetes-native nature and the support for all the roadmap items we already had: VPA, automatic node sizing, partial Spot utilization.
Ritwik G, Staff Engineer at Unstract
Onboarding began with the staging cluster. Ritwik was initially cautious – handing control of the workload’s CPU and memory to an automated system felt like a significant step – and the team spent roughly a month in staging before promoting the configuration to production.
Initially, there was a learning curve, but of course, the onboarding team was patient and helped us throughout the process.
Ritwik G, Staff Engineer at Unstract
Unstract’s confidence grew as the observed behavior matched the explanation. Cast sets memory requests close to actual usage patterns (derived from rolling history) while leaving limits with headroom for spikes – a subtle but important departure from Unstract’s prior practice of equating requests and limits. The system also automatically increased memory when a workload approached its limit, without any human action.
Production cluster monthly cost before and after Cast AI
Today, Unstract runs Cast across all production clusters in largely autonomous mode. Scheduled rebalances optimize spot utilization on interruptible workloads. Cast’s recommendations serve as the baseline for any remaining manual tuning. The team reviews the platform monthly rather than daily.
Nowadays, it is almost in an autonomous mode. We are rarely touching things. I’d say at least 80% of our time spent on DevOps is now cleaner – a lot of those Kubernetes-related things just happen.
Ritwik G, Staff Engineer at Unstract
Overall cost savings for three clusters
The Results: Reliability First, Cost Savings as a Bonus
- 80% reduction in the time spent on Kubernetes operational work
- Memory-related alerts downgraded in severity (P0 → P1) as automated remediation made them non-critical
- 30% reduction in production infrastructure costs since adopting Cast
- 23% overall cloud cost reduction across all environments
Massive time savings
The shift from reactive to continuous optimization has also removed a chronic source of rework. Infrastructure rightsizing is no longer a periodic project that decays between cycles – it’s a background process that automatically keeps pace with Unstract’s evolving workloads.
Reliability boost
The reliability impact has been as significant as the cost savings. OOMKill-related incidents – once a source of nightly on-call escalations – have effectively disappeared. Memory alerts that previously required immediate human response have been reclassified from P0 to P1 because Cast automatically remediates them before the situation becomes critical.
Reliability has improved significantly. Before Cast AI, critical components like Redis could experience prolonged downtime when memory spiked – whether from an unusual request pattern or a problematic file in the pipeline. We always kept memory requests and limits identical, pegged to the highest expected usage plus a buffer.
What I observed with Cast is that it keeps the request close to actual typical usage, with the limit a bit higher. So even if a rogue request comes in, it can spike and come back down. Memory is set based on three days of actual usage history rather than load tests – and load tests are never a one-to-one representation of real traffic.
Because of this, we actually downgraded our memory alerts from P0 to P1 – when an alert fires at 90% utilization, Cast has already handled it before it becomes critical. It’s been a real blessing for the on-call folks; they don’t need to wake up at night the way they used to.
Ritwik G, Staff Engineer at Unstract
Full Kubernetes automation
Perhaps the most telling outcome is how the team now thinks about Cast. What began as a cost optimization tool was quickly reframed as something more fundamental.
While we started exploring Cast as a cost optimization tool, we soon realized that the cost optimization was a side effect of DevOps optimizations. The platform has also helped us eliminate some of our planned DevOps automation work, as these capabilities were already readily available within Cast AI. The VPA feature, in particular, has been extremely helpful in improving the platform’s overall stability, reducing many routine on-call alerts, and easing the team’s fatigue.
Many of the DevOps tickets we had sitting in our backlog were automatically covered once we onboarded. This way, our team can focus more on the application side of things.
Ritwik G, Staff Engineer at Unstract
Looking ahead, Ritwik sees Cast’s expanding feature set as a pipeline of DevOps improvements the team can adopt immediately, without having to research, learn, and implement them from scratch.
We believe Cast will bring even more DevOps automations, which we can immediately leverage, rather than our team needing to research, learn, and implement these. We focus on application deliverables. This is a perfect fit for us.
Ritwik G, Staff Engineer at Unstract
