AI prototype to production: turning a demo into a stable system with an SLA
Taking an AI prototype to production when it shines in demos but fails under real load. We audit, harden and launch it on part of traffic, then keep it running under an SLA.
What hurts
The AI agent prototype was built in a few weeks, and the demo works. The road from prototype to production turns out longer. With real customers the agent gives odd answers, and no log shows why.
The code was written for speed: no tests, keys in the repository, manual deploys from a laptop. Every prompt or model update breaks what worked yesterday.
Your security team or a large customer asks for an SLA, an access log and an incident response plan. None of it exists, and the launch has already been promised.
What we build
- Audit of code, architecture, security and AI behaviour on real data
- A prioritised remediation plan with a fixed estimate
- Refactoring of critical parts and removing secrets from the code
- An AI test suite and automated tests for the key flows
- CI/CD: automated builds, checks and safe deployments
- Monitoring and alerts on Prometheus and Grafana, a log of AI answers
- Launch on part of traffic with fast rollback
- Documentation and an incident response plan
What you get
- A stable system in production with configured deployments
- An audit report listing risks found and what was fixed
- A test suite that runs before every release
- Monitoring dashboards and alerts
- Documentation and an SLA support contract
How we measure the result
- Share of AI test cases that pass before a release
- Production incidents per month
- Time from detecting a problem to the fix
- Share of releases without a rollback
Systems and stack
Industries
Questions
How much does it cost to take a prototype to production?
From $10,000. The price depends on the state of the code, the number of integrations, security and load requirements. We start with an audit, then give a fixed estimate within 2-3 days. If we overrun it through our fault, we finish at our own cost.
How long does it take?
Usually 6-12 weeks. The first weeks cover the audit and the most urgent security fixes. Then tests, monitoring and a launch on part of traffic. You see what changed every week.
Will you rewrite everything?
Not if we can avoid it. We keep what works and rewrite only what blocks stability. If the audit shows a rewrite is cheaper, we will say so plainly and explain why.
Another team built the prototype. Is that a problem?
No. We need access to the code and environments, and ideally a short call with the authors. We sign an NDA before access. After the audit you keep the documentation whoever works on the system next.
How do you test AI when its answers vary?
We build test cases from real conversations and tasks, and check the meaning of the answer rather than exact wording. We do the same on our own Happ platform. AI sometimes gets things wrong, so besides tests there is monitoring and a hand-off to a person for hard cases.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.