From an audit that said "not ready" to an AI platform in production.

Theenk.AI is Olensys's own product, and by mid-2026 it was no greenfield prototype: 12 repositories — a core API, an LLM service, a scraper, a serverless orchestrator, an analytics pipeline, a dashboard, an embeddable widget, a public site, a self-hosted inference gateway and observability stack, infrastructure, docs — built fast, with three to seven engineers touching each.

Four facts on record set the starting position. There was no production environment: everything ran in dev. No secret was under a manifest check. The agent evals were six numeric "quality" scores. Billing ran on Chargebee and was not live.

On 2 July 2026 Olensys ran a production-readiness audit across all 12: nine parallel read-only reviews, every finding tied to a file and line. Its verdict: not launch-ready. The application code was in better shape than the infrastructure around it. This page gives the audit's verdict, not its findings.

For the business, that meant no launch: without a production environment, there was nothing to put operators on.

Assessment and engineering plan. The remediation plan separated what could be fixed in code from what needed production infrastructure. Code came first, in two days: on 3 July, seven repositories changed. Every one built, linted, and type-checked clean, with verbatim output captured. Nothing was committed until a human read it. The release plan, written 31 Aug, then made infrastructure the gate.

AI evaluation. The agent evaluation suite was rebuilt from six numeric "quality" scores into 15 binary evaluators — five deterministic, 10 managed LLM judges — each targeting a failure mode that had actually been observed. Three datasets (sales, support, productivity) replay real user turns through the live agent; every item produces a trace in Langfuse. A pass is evidence about those cases, not a measure of how the agent behaves in production. Below is the evaluation architecture.

Production architecture. There was no prod environment in Terraform. Olensys built the production environment from nothing: networking, compute, identity and CI access.

Reliability and controls. A manifest catalogs 229 secret parameters across dev and prod, and a CI step checks them on every infrastructure PR. Values never printed. Nine orphans deleted. The delivery system Olensys sells runs here: spec, plan, verification, and a named engineer on every merge. Agents open draft PRs with evidence; a hook stops any agent from merging, tagging, or pushing to main. Model routing is written down: which model for which task, and why. Session handoffs are files, not memory. Below is the production engineering workflow.

Production deployment. The services went live wave by wave, in dependency order, by early September 2026. Billing moved from Chargebee to Paddle and went to production with the platform.

• What failed: a CI runner ran out of disk, twice, over three weeks. Why it failed: a CUDA build of PyTorch (about 3 GB of GPU wheels) was being installed on a CPU-only test runner, with a cache that never restored. What changed: one line, to install from the CPU index. The full disk had been masking a red pipeline since early September.
• What failed: a core API's task definition was x86 on an ARM cluster. Why it failed: it worked in dev by accident; it would not have worked in prod. What changed: the production build pins the core API to ARM64.
• What failed: the first eval suite measured the wrong thing. Why it failed: six numeric scores produced numbers nobody could act on. What changed: 15 binary evaluators, each tied to a failure mode that had been observed, replaced them.

The audit of 2 July 2026 said not launch-ready. By early September 2026 every service was live in production. The table sets the platform on the day of the audit (Before) beside October 2026 (After).

The rows count engineering work and where it landed, not production reliability. "Agent evaluators" counts the checks the suite runs, not how often the agent passes them. "Secrets under manifest check" counts the secret parameters the infrastructure CI step checks, across dev and prod. The commits and merged pull requests are across the 10 service repositories, as of October 2026.

Methodology: all figures from the repositories, Terraform state, the production deploy record, and the 2 July 2026 audit. This is Olensys's own system, instrumented the way client systems are. It is not a client result.

Where the capacity went

Here is where the effort went, by what each part delivered. The code pass: seven repositories changed in two days. Evaluation: 15 pass/fail evaluators over three datasets of real user turns. Architecture: a production environment built from nothing. Reliability and controls: the secrets manifest checked on every infrastructure PR, and a hook that stops any agent from merging, tagging, or pushing to main. The release: the services went live wave by wave, in dependency order.

The same team that would have spent the fall on a production build-out spent it on the product: outcome-driven agent behavior, template capabilities, a Slack pre-sales agent, model-catalog tiers.

Not every audit finding is closed. The open ones are dated and tracked on the backlog.