Production reliability
for AI systems
An AI system in production is more than its model. It's pipelines, networks, permissions, data, monitoring and cost, all changing at once, and the failures that stall AI projects usually live in how those parts fit together. We look for where a system is fighting itself and give it a structure that adapts instead of breaks.
Crucible Science is an independent research lab. Our current work is on how AI systems behave in production, and we work hands-on with teams to apply what we find.
- Launch readiness
- Your pilot works, and now you need evidence it's ready for real users: evaluation and release gates in your pipeline, so changes ship on proof.
- Boundaries
- What your agents and services can reach, defined in code, with approval paths for the actions that matter.
- Change and rollback
- Recovery you've rehearsed, and a plan for when your provider updates or retires the model you depend on.
- Observability and cost
- Knowing what your system actually did, why, and what each finished task costs.
Engagements are shaped to the work: a fixed-scope project, a retainer, or covering a role until you hire, remote or on site.
- Note three What changes in your agent when its model is upgraded? Claude Code, Opus 5 to Opus 5.5.
- Note one What repeated runs of a customer-service agent show that a single pass rate hides.
- Note two One coding agent across five models, and one model in five harnesses, and what moved.
- CheckAgain A free, open-source page that reads repeated runs before and after a change. Your file never leaves the browser.
Ashwin founded Crucible Science after more than eight years in engineering across electronics, cloud and AI, including five years at a Big Four consultancy, in the core team that moved cloud operations from manual work to automated code deployments across more than 40 Azure subscriptions.
- Built deployment pipelines with Terraform, Ansible, Azure DevOps and Kubernetes. One master pipeline stood up five to ten applications, with security and networking, in about two hours instead of weeks.
- Led network design in code: virtual networks, subnets, security rules, routes and firewall policy, which together set the boundaries of what every system could reach.
- Led observability across the fleet. Ran failure drills that destroyed environments and rebuilt them from code, so recovery was rehearsed rather than hoped for.
He holds an M.S. in Technology Systems Management (Data Science) from Stony Brook University. Before it, he founded a small bootstrapped venture that was profitable in its first year.
He has filed a patent for an adaptive augmented-reality system, taking it from idea to filed specification in six months.
Tell us what you're running and what's worrying you.
ashwin@cruciblescience.com