Your AI feature holds in production.

For funded B2B SaaS teams whose shipped AI feature hallucinates or breaks. We make one named feature production-reliable — adversarial evals, guardrails, observability — and then operate it, so it keeps holding as the models and your product change and your engineers stay on the roadmap.

The reply is signed by the person accountable for the work. No form, no call widget.

What a teardown finds

We use your feature the way a careful new customer would — public surface, free tier, or a trial we sign up for as ourselves — and run a fixed battery of adversarial inputs: features that do not exist, plan limits with a known ground truth, false premises, empty and whitespace input, out-of-range identifiers, prompt injection, and the same question asked three times in fresh sessions. Every failure we report is reproduced at least twice before it goes in the report.

Method, in three lines: read-side only — we never touch production secrets, other users' data, or PII; human pace, no load, no scraping; every input and response is quoted verbatim in the report so your team can reproduce it in minutes.

See the report format (sample — run against a local mock, not a customer)

How it works

1. Free teardown. You email us the feature; we confirm a date. You get a private, unlisted report: the exact inputs, what happened, the failure classes, and a one-page fix plan. You judge the result, not a pitch. Cost: nothing.
2. Pilot — $4,000, about two weeks. We fix one failure class end to end and instrument it: an eval suite for that class (delivered to you, runnable by you), the guardrail or grounding change, and tracing so the class is caught before release instead of by users. 50% on signature, 50% on delivery.
3. Audit + fix — $8,000. One named feature made production-grade: the full adversarial suite, the guardrails (input/output validation, retrieval-grounding checks, low-confidence fallback and human handoff), and observability wired in.
4. Operate — $3,000 per month, month to month. Offered after the pilot or audit has shown its result. The suite runs on every change and on a schedule, production traces are watched with quality alerts, you get a monthly regression report, and we iterate as models and product move. This is the part a tool cannot do for you.

Fixed prices

All prices in USD. No open-ended scope; every engagement names the feature and the eval suite up front.

Who this is for — and not for

For

• Funded seed to Series B B2B SaaS with a shipped AI feature you own: your pipeline, your prompts, your retrieval, your tool calls.
• Support copilots, RAG assistants, agentic workflows, in-app "ask AI" — anything whose failures reach your customers.
• Teams that would rather keep their engineers on the roadmap than have them babysit the agent.

Not for

• Teams whose only AI surface is a vendor's docs widget — you cannot change the pipeline, so there is nothing for us to fix.
• Pre-launch prototypes, and consumer apps.

Questions, answered plainly

Do you need production access?

No. The teardown is read-side on the public surface or a trial seat. For the pilot we work from a staging key or a trial seat; production credentials are never a precondition.

Do you keep our data?

No. The inputs we send and the responses we get go into the report we hand you and nowhere else — we never publish a named finding. We do not scrape, store or reuse your customers' data.

Why not just run Braintrust, Arize or Galileo?

You can, and we will use whatever tooling you already have. A tool gives your team one more thing to operate: someone still has to write the adversarial cases, decide what a failure is, watch the traces and act on regressions every week. We operate it.

Who does the work?

Holds in Prod is run by Finn Wiechert, who is accountable for every report and every fix and is who you deal with. No account manager, no hand-off to a team you have not met.

What if it does not work?

The operate retainer's first month is refundable if the failure rate on the eval suite agreed and frozen at pilot start does not measurably drop. The pilot has a fixed scope and a fixed price; you see the eval suite before you pay the second half.

Do you work with EU companies?

Yes, on inbound. We never cold-email EU addresses.

How long until the first result?

We tell you the date when we confirm your email; a pilot is about two weeks from signature.

Who we are

Holds in Prod is run by Finn Wiechert, an independent engineer in Germany, and works on one thing: shipped AI features that have to hold in production. Every report and every fix goes out under his name and accountability. We do not publish named teardowns; what we find about your product goes to you.