Work · Dynamic CX

One AI-built platform, taken to the top tier

Dynamic CX built a contact-centre intelligence platform with AI, put it in front of real customers, and in September 2025 asked us to make it as strong as it is fast. What arrived, what was done, what changed.

What arrived

A founder-built platform, moving fast, already in production.

Dynamic CX is contact-centre intelligence: calls in, transcripts and AI analysis out, the CRM updated without anyone typing. Its founder built it with AI, quicker than a dev team would have, and real tenants were using it. What it did not yet have was the ground under it: hardened environments, automated releases, the security posture buyers ask about. The brief was blunt: keep the speed, and take the platform to the standard buyers, investors and enterprise customers expect.

What was done

The ground first, then a system that keeps speed safe.

The foundations came first. One hand-run environment became development, staging and production, defined in code: automated deployments, separated deployment and runtime identities, managed secrets, and verification steps that fail a release before it can fail a customer. On that ground, a review system: plan-first reviews, every pull request read before staging, and each finding folded into checks the AI tooling applies from then on. Alongside review, we led builds inside the platform: the pipelines that carry every call from ingestion to analysis, an AI matching engine on Claude and Gemini, and the integrations that write results back into the tools its customers already run on.

What changed

Faster than before, and safe at that speed.

The platform did not slow down to get safer. More change landed in the last seven weeks than in its first three months, with more than 15,000 automated tests now standing behind it. High-severity issues are caught before staging rather than by customers; release checks cover every deployment, and anything they flag is fixed the same day. In the second week a background job went from 55 seconds to 2. Nearly a year on, the engagement is still running: every pull request reviewed, releases landing week in and week out, and a plain-English note on what was done and what is next.

What top tier means in practice

  • Infrastructure as code

    Development, staging and production are deployed from version control, with deployment and runtime identities separated, secrets managed centrally, and pre-flight checks that block a bad release. Twenty-eight releases shipped in the first six weeks; when a hosting fault failed twelve consecutive deployments, it was debugged and fixed with zero downtime. A full restore has been proven for real, not just tested: 223 files back within hours.

  • Pipelines built for failure

    Every call moves through independent scheduled jobs: ingestion, transcription, analysis, CRM posting. Each step is idempotent, retried with backoff and capped so failures cannot compound, and claim-based locking stops parallel workers paying to process the same call twice. Fallback sweeps catch anything missed. A surge of calls slows nothing down.

  • Review as a system

    More than 900 pull requests reviewed since October 2025, with plans reviewed before code is written so architectural problems are settled on paper. Catches before staging include cross-tenant data leaks, SQL and command injection, prompt injection and OAuth flaws. Every finding becomes a check the AI tooling applies on future changes, which is why issues per pull request keep falling.

  • Determinism where it matters

    Where AI decides, rules constrain: a deterministic layer enforces hard business rules over model output. Where money moves, writes are transactional and carry idempotency keys, so a retry can never charge twice. Where load is unknown, queries are bounded and memory use is measured rather than guessed. The platform behaves the same way every time; that is what makes it sellable.

The numbers

55sto2s
a background job, in the second week
1,051to2
database calls per cycle, same week
900+
pull requests reviewed across the engagement
200+
high-severity issues caught before staging, one five-week period
28
releases in the first six weeks, once deployments were automated
15,000+
automated tests behind the platform

Next

Ready to take your platform to the top tier?