# From Experiment to Production What production needs beyond an experiment, and how Perceivo keeps compute, services, storage and secrets together. - Page: https://perceivo.ai/university/infra-01-experiment-to-production - JSON: https://perceivo.ai/registry/film/infra-01-experiment-to-production.json - Kind: film - Revision: 3 (published 2026-10-18, produced 2026-10-15) - Previous revision 1 (published 2026-10-07): https://perceivo.ai/registry/film/infra-01-experiment-to-production/r1.json - Previous revision 2 (published 2026-10-11): https://perceivo.ai/registry/film/infra-01-experiment-to-production/r2.json - SDK version: none ## Validation - SDK: not applicable (Concept film; no component or orchestration is offered for download) - Platform: not applicable (Concept film; no component or orchestration is offered for download) ## Transcript [0:00] Introduction to Infrastructure, Episode 1: From Experiment to Production [0:06] Welcome to our Introduction to Infrastructure series. [0:09] This series is about where your work runs, and the tools it has to work with. [0:15] Let's start with an experiment. [0:18] In Perceivo, an experiment is where an idea gets tried: different prompts, different models, runs side by side, until one of them wins. [0:28] Then production asks for more. [0:31] The work has to run on compute that grows and shrinks with demand. [0:36] It needs somewhere to keep memory, and data. [0:39] It needs credentials it can use without anyone pasting them into code. [0:44] And someone has to be able to see what's running. [0:48] Each of those is usually its own project. [0:51] That's a large part of what makes the step to production expensive. [0:56] Perceivo's infrastructure was built to make that step smaller. [1:01] Compute, services, storage and secrets sit in the same place as your experiments, your orchestrations and your intelligence services. [1:11] So the experiment that won already runs where production runs. [1:16] Each organization gets its own namespace. [1:19] Think of it as a fenced-off space, just for you. [1:24] Inside it, services: Ray, Redis, Milvus, Elasticsearch and an MCP server. [1:32] And workers: training, inference, evaluation, features and orchestrations. [1:39] All sharing your compute limit. [1:42] Files and service data live in your organization's own storage container. [1:47] And secrets are kept in Azure Key Vault. [1:51] The namespace, the services, the workers and the storage belong to your organization. [1:58] Perceivo tracks compute usage in compute units, which count processing and memory as one joint unit. [2:07] One compute unit is one hour of runtime with one virtual CPU and four gigabytes of memory. [2:14] And it scales up or down with what is actually consumed. [2:18] The compute in your subscription plan runs on the shared, default nodepool. [2:24] Workloads can be run one-shot, or they can be persisted. [2:28] When you run an orchestration one-shot, it is truly on-demand compute. [2:34] It starts, runs once, and when it finishes, your compute, and your billing, go back to sleep. [2:41] If you have to run orchestrations continuously, such as when it is driving a website chatbot, you may want to keep multiple replicas available around the clock. [2:52] Larger plans provide more compute, but some customers may also benefit from compute that isn't shared. [2:59] So you can also provision a dedicated nodepool: CPU or GPU nodes, that only your organization uses. [3:07] A nodepool is created with one simple dialog. [3:11] You are in control of the resources allocated to every resource and workload in your namespace, including which nodepool it runs on. [3:21] Compute runs the work. [3:23] But AI solutions need more than compute. [3:26] They need memory, data, credentials, and models. [3:31] So the platform provides tools, in two forms. [3:35] Some run inside your namespace, started when you need them, and stopped when you don't. [3:41] Redis, for fast, short-term memory. [3:44] Milvus, a vector database, for high-dimensional vector embeddings. [3:49] Elasticsearch, for high-performance long-term memory. [3:53] And an MCP server, for structured access to your enterprise data. [3:59] Each of them backs itself up to blob storage, so they all survive a restart. [4:06] Others are managed services, run in Azure for your organization. [4:10] Azure Database for PostgreSQL, for relational data. [4:15] And Azure Cache for Redis, for high throughput and redundancy. [4:21] Secrets sit alongside them, in Azure Key Vault. [4:24] You create a secret once, and bind it to each component that declares it needs one. [4:30] The orchestration refers to it by name. [4:34] You can host a model in your namespace if you like, but you have access to over 11,000 enterprise AI models through the Perceivo SDK. [4:44] Orchestrations reach these platform services in two ways. [4:48] First, Perceivo provides prebuilt system connections as templates, that you add and configure on the canvas. [4:57] Or you can create a custom implementation, and access platform services through the SDK. [5:04] The Infrastructure overview gives you a snapshot of the overall condition of your infrastructure: running pods, real-time metrics, consumption and cost. [5:14] It shows how much of your capacity is in use right now, split between your own workloads, and the platform services that support them. [5:23] How that use has changed over time. [5:25] Which nodepools and pods are running. [5:28] What's waiting in the queues. [5:30] Which services are up. [5:31] And how your storage is used. [5:34] It refreshes every fifteen seconds, and its panels link to the page that controls each resource. [5:42] From here, the series takes one feature at a time. [5:46] Compute and the Ray cluster. [5:49] Dedicated nodepools. [5:51] Redis, Milvus, and Elasticsearch. [5:55] Managed PostgreSQL. [5:57] Secrets. [5:58] The MCP server. [5:59] Hosted models. [6:01] Storage. [6:02] And running it all, day to day. [6:06] Next time, compute and the Ray cluster. [6:10] See you there.